Artiﬁcial Intelligence as a Hope

Volume 65 Number 6 July 2016

Artificial Intelligence as a HopeAI for Taking on the Challenges of an Unpredictable Era

www.hitachi.com/rev

From the Editor

The term artificial intelligence (AI) has started to become

commonplace in Japanese print media and TV. However, since

many people are still largely unfamiliar with what the term actually

means, my office receives questions about AI and requests for

presentations about it on a daily basis.

AI has certainly been covered in a large number of books and

other media, however only certain aspects of it tend to be

presented. As the term AI is currently trending, some of the

coverage has started to exploit the trend by interpreting it in a

broad sense, giving people the wrong impression.

Another problem is that the latest developments in AI are taking

place at centers of corporate activity, so they are restricted by

corporate confidentiality, making mass media exposure difficult.

At the same time, only the technologies that can be shown to the

general public, such as web-based image recognition, are being

covered as examples of AI, creating a poor balance of coverage.

This feature issue attempts to go beyond the current coverage to

present a total picture of the many different aspects of AI. The first

half contains several views of AI presented by National Institute of

Informatics professor, Noriko Arai, and Yahoo Japan Corporation

Chief Strategy Officer, Kazuto Ataka. The second half looks at the

breadth of AI applications and depth of the technology involved,

along with the innovations that are making them possible using the

same general-purpose AI. There is information on AI applications

and technology that have expanded into industries such as finance,

railway, distribution, water and manufacturing.

Many readers will doubtlessly view a collection of corporate

articles like this one as a form of corporate PR. However, while this

issue uses Hitachi examples to describe AI, it is designed as a

special feature that will satisfy the interests of many readers who

are eager to learn about the current state of AI. We have aimed to

make this issue something that busy business people will be willing

to pay to read. Whether we have succeeded in this aim will be up to

our readers to determine.

This April, Hitachi strengthened its front-office organization,

making a fresh start with a new organization for creating

innovations in collaboration with customers. We are showcasing AI

as the core technology for this new approach.

Japan is now at the turning point of a shift from a

manufacturing-based economy to an economy based more on value

generated from services. However, there is a need to survive this

age of innovation as the effects of never-ending global economic

changes and conflicts extend to each and every one of us. I think

that AI provides hope for survival in this unpredictable age. I hope

this issue will help readers take advantage of this new ray of hope.Kazuo Yano, Dr. Eng.Corporate Officer

Corporate Chief Scientist,

Research & Development Group,

Hitachi, Ltd.

Editorial Coordinator,

“Artificial Intelligence as a Hope:

AI for Taking on the Challenges of an

Unpredictable Era” Issue

Artifi cial Intelligence as a Hope AI for Taking on the Challenges of an Unpredictable Era

ContentsExpert Insights

6 Approximate Solutions and True SolutionsNoriko Arai

Technotalk 8 Human-friendly AI that Learns from Life

Kazuto Ataka, Kazuo Yano

Overview 14 AI for Taking on the Challenges of an Unpredictable Era

Kazuo Yano

Featured Articles 35 AI Technology

Achieving General-Purpose AI that Can Learn and Make Decisions for ItselfNorihiko Moriwaki, Tomoaki Akitomi, Fumiya Kudo, Ryuji Mine, Toshio Moriya, Kazuo Yano

40 AI Services and Platforms

A Practical Approach to Increasing Business SophisticationYasuharu Namba, Jun Yoshida, Kazuaki Tokunaga, Takuya Haraguchi

45 Utilization of AI in the Financial Sector

Case Study and Outlook for FinTech EraKiyoshi Kumagai, Satomi Tsuji, Hisanaga Omori

50 Utilization of AI in the Railway Sector

Case Study of Energy Efficiency in Railway OperationsRyo Furutani, Fumiya Kudo, Norihiko Moriwaki

56 Use of AI in the Logistics Sector

Case Study of Improving Productivity in Warehouse WorkJunichi Hirayama, Tomoaki Akitomi, Fumiya Kudo, Atsushi Miyamoto, Ryuji Mine

61 Utilization of AI in the Water Sector

Case Study of Converting Operating History Data to ValuesIchiro Embutsu, Koji Kageyama, Satomi Tsuji, Norihiko Moriwaki, Yukiko Ichige

67 Utilization of AI in the Manufacturing Sector

Case Studies and Outlook for Linked FactoriesNaohiko Irie, Hiroto Nagayoshi, Hikaru Koyama

73 Advanced Research into AI

Debating Artifi cial IntelligenceKohsuke Yanai, Yoshiyuki Kobayashi, Misa Sato, Toshihiko Yanase, Toshinori Miyoshi, Yoshiki Niwa, Hisashi Ikeda

78 Advanced Research into AI

Ising ComputerMasanao Yamaoka, Chihiro Yoshimura, Masato Hayashi, Takuya Okuyama, Hidetaka Aoki, Hiroyuki Mizuno

Artificial Intelligence as a HopeAI for Taking on the Challenges of an Unpredictable EraAAAAAAAAAAAAAAAAAAAAAAAAAIIIIIIIIII fffffffffffffffffffoooooooooooorrrrrrrrrrrrrrrrrrrrr TTTTTTTTTTTTTTTTTaaaaaaaaaaaaaaaaakkkkkkkkkkkkkkkkkkkiiiiiiiiiiiinnnnnnnnnnnnnggggggggggggggggg ooooooooooooooooonnnnnnnnnnnnnnnn tttttttttttttttthhhhhhhhhhhhhhhhhhhhheeeeeeeeeeeeeeeeee CCCCCCCCCCCChhhhhhhhhhhhhhhhhhhaaaaaaaaaaaaaallllllllllllllllllllllllllleeeeeeeeeeeeeeennnnnnnnnnnnnngggggggggggggggggeeeeeeeeeeeeeeessssssssssssssssss ooooooooooooooooofffffffffffffffffff aaaaaaaaaaaaaaaaannnnnnnnnnnnnn UUUUUUUUUUUUUUUnnnnnnnnnnnnnpppppppppppppppppppprrrrrrrrrrrrrreeeeeeeeeeeeeeeeeeddddddddddddddddddddddiiiiiiiiiiiicccccccccccttttttttttttttaaaaaaaaaaaabbbbbbbbbbbbbbblllllllllllllllleeeeeeeeeeeeeeeeeeee EEEEEEEEEEEErrrrrrrrrrrraaaaaaaaaaaaaaaaaArtificial Intelligence as a HopeAI for Taking on the Challenges of an Unpredictable Era

Artificial intelligence (AI) has started to move beyond the realm of computer science and find genuine practical uses.What impact this will have on business and other parts of society is a matter of growing interest.A general-purpose AI developed by Hitachi is being put to use in a wide range of industries where it is successfully using past data to generate knowledge for a better future and to devise accurate ways to deal with the problem of unknowns.In doing so, it is giving a foretaste of new ways of confronting the unknown.AI is beginning to demonstrate its great potential as a source of hope for us, living in unpredictable times of relentless change.

AAAArArArArArArArArArArArArArArArArttttittitititititititiifififififififififififif ccciciciciciciciciciciciialalalalalalalalalalalalala iiiiiiiiintntntntntntntntntntntntteeleleelelelelelelelelellililiiililililililigegegegegegegegeggggggg ncncncncncncncncnce e e e e e e e e eeeeee (A(A(A(A(A(A(A(A(A(A(A(A(AA( I)I)I)I)I)I)I)I)I)I)I)I)I)I) hhhhhhhhhhhhhhhhhhasasasasasassasasasasasass ssssssssssssssstatatatatatatatatatatatatatartrtrtrtrtrtrtrtrtrtrtrtrtrtrttedddddededededededededededededede ttttttttttttttooooooooooooo momomomomomomomomomommomoovevevevevevevevevevevevevee bbbbbbbbbbbbbbbbbbeyeyeyeyeyeyeyeyeyeeyeyeyyyyyyyyononononononononononoondddddddddddddddddd ththththththththththththhhththththe e ee e e e e eee eeee rerererererererererereeaaalalalalaalalalalalalalmmmmmmmmmmmmm ofofofofofofofofofofofofofofofofofof cccccccccccccomomomomomomomomommpupupupupupupupupupppppppppp tteteteteteteteteteteteteter rr r rrrrrr r r r r r r r r scscscscscscscscscscscscscscss ieieieieieeieieieieieieiencncncncncncncncnce e ee e e e e eee eeee aaananananananaananananana dddddddddddddddddd fffififififififififififindndnddddndndndndndndndndndndndnd gegegegegegegegegegegegegegeenununununununununuinininininininininini e e e ee e e e e e eeeeee prprprprprprprprprprprpppppppp acacacacacacacacacaca ttitititititititititicacacacacacacacacaal ll lll lllll ususususususususususussusu esesesesesesseseseseseseses........WhWhWhWhWhWhWhWhWhWhWhhhWhWhWhWhWh ttatatatatatatatatatatatatatatat iiiiiiiiimpmpmpmpmpmpmpmpmpmpmpmppppppppacacacacacacacacacaca tt t t t t t tttttt ththththththththhhthththththisissssissisisisisisisis wwwwwwwwwwwwwwwwwwiiliilililililillllllllllll hahahahahahahahhhhhahahahahahahaavevevevevevevevevevevevevee ooooooooooooooonnnnnnnnnn bubububbbbbbbubububububububu iisisisisisisisisisineneneneneneneneneesssssssssssssssssssssssss aaaaaaaaaaaaaa ddddddndndndndndndndndndndndnd oooooooooooooooththththththththththhhhthththththererererrererererererrerrrrrr pppppppppppppppppppararararararararararara ttststststststststststststss ooooooooooooooof f f f f f f fffffff f f f ssosososososososososososos iciciciciicicicicicicicii ttttttetetetetetetetetete yyyyyy y y y yyyy y y y y y yy isisisisisssissisisisisisiss aaaaaaaaaaaaaa mmmmmmmmmmmm tttttatatatatatatatatattetttttteteteteteteteter r r r r rrrrrr r r r r r r r ofofofofofofofofofofofofofofofofofof ggggggggggggggrororororororororororororowwwwiwiwiwiwiwiwiwiwiwingngngngngngngngngngngngnggg iiiiiiiiiii ttttttntntntntntntntntnttererererererererererrresesesesesesesesesesese tttt.t.tt.t.t.t.t.tttAAAAAAA A A AAA AAAAAAAA gegegegegegegegegegegegegegegegegegeeneneneneneneneerarararararararararaal-l----l-l-l--l-l-l-l-l pupupupupupupupupupupupupupupupppp rprprprprprprprprprprprprppppppposososososososososse ee e eee e e eeeeeee AAAAAAAAAIAIAIAIAIAIAIAIAIAIA dddddddddddddddddddevevevevevevveveveveve eelelelelelelelelelelelopopopopopopopopopopopopoppppppp dddddddededededededededededede bbbbbbbbbbbbbbbbbyyyyyy yy yyyy y y y y y y yy HHHHiHiHiHiHiHiHiHiHitatatatattttatatatata hhhhhhhhchchchchchchchchiiiiiiiiii isissssssisisisisisisisiss bbbbbbbbbbbbbbbbbb ieieieieieieieieieieingngngngngngngngngngngngngng ppppppppppppppppppp tttututututututututututututut ttttttttttttto o o o oo o o o ooooooo ususususususususususususse e e e e e e e eeeeeee ininininininininininin aaaaaaaaaaaaaaa wwwwwwwwwwwwwwwwididdddddidididididididideeeeeeeeeeee rarararararararararaaangngngngngngngngngngngngngggeeeeeeeeeeeeee ofofofofofofofofofofofofofofofofof iiiiiiiii dddddddddddndndndndndndndususususususususstttttrtrtrtrtrtrtrtrtriiieieieieieieieiees s s s ss sssssssss

hhhwhwhwhwhwhwhwhwhwhwhwhwhwhwhwhw ererererererereererererreeeeeeeeeeeee ititititititittititititit iiiiiiiiiissssssssssssss susususususususususususususs ccccccccccccccccccccesesesesesesesesesessfsfsfsfsfsfsfsfsfsfsfsfsfs uluuuuululululululululllylylylylylylylylylylylylyylylylylylyy uuuuuuuuuuusisisisisisisisisisisisingngngngngngngngnggggg ppppppppppppppppppasasasasasasasasasast t t t t t t t tttttt ddddadadadadadadadadadadadadadadadaatatatatatatatatatatatatatat ttttttttttttttoooooooooooooo gegegegegegegegegegegegegegeegg neneneneneneneenerarararararaararararateteteteteteeetetetetetetetete kkkkkkkkkkkkkkkkknononononononononononowlwlwlwlwlwlwlwlwlwlw ededededddddedededededededededede gegegegegegegegegegegegegeege ffffffffffffforororororororororororororrorrrr aaaaaaaaaaaaa bbbbbbbbbbbbbbbbbbeeetetetetetetetetetetetete tetetetettteteteteteteteter rr r rrrrr rr r rr r rr r fufufufufufufufufufufufufufufuttttututututututututututurererererererererererererereree aaaaaaaaaaaaaaaa dddddndndndndndndndndndndndndnd ttttttttttttttoooooo o o o o o o o oooo ddddedededededededededededededededeevviviviviviviviviviseseseseseseseseseeseseseseseaacacacacacacacacacacacaca cucucucucucucucucucurarararararaarararararateteteteeeetetetetetetetetetete wwwwwwwwwwwwwwwwwwayayayayayayayayayaayayyyyyyyyyssss s s s s sss ssss tototototototototototototototototot dddddddddddddddddddeaeaeaeaeaeaeaeaeaealllllllllll wwwwwwwiwiwiwiwwiwiwiwiwiwiwithththhhhhhhhhhththhththththt ttttttttttttttthehehehehehehehehehehhhhehehehehehee pppppppppppppppppprororororororororororororoblblblblbbbbblblblblblblbllemememememememmmemememem ooooooooooooooof f f f f f f f f f f ffffff ununununununununununuu kknknknkkkkkkkknknknknknknknk owowowowowowowowowowwowownsnsnsnsnsssnsnsnsnsnsnss.........InInnnInInInInInInInIn ddddddddddddddddddd iioioioioioioioioioioiingngngngngngngngngngngngngg sssssssssssssso,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,oo iiiiiiiiit t t t ttt ttttttt isisissssissisisisisisisiss ggggggggggggggivivvviviviviviviviviviinininininininininininggggggggggggg a a a a a a a a aaaaaaaa fffofofofofofofofofofofofofoforererererererererereereettttttatatatatatatatatatata ttttstststststststststststeeeeeeeeeeeeeee ofofofofofofofofofofofofofofofofof nnnnnnnnnnewewewewewewewewewewewewewwwww wwwwwwwwwwwwwwwwayayayayayayayayayayayayyyyyyyys ss s s s s ssssssss ofofofofofofofofofofofofofofofofofof ccccccccccccccoononononononononononfrfrfrfrfrfrrfrfrfrfrfrfrfrononononononononononoo ttitititititititititititingngngngngngngngngngngnggg tttttttttttttthehehehehehehehehehhhhhehehehehehee uuuuuuuuuuuuu kkkkkkkkknknknknknknknknknkknonononoonononononononownwnwnwnwnwnwnwnwnwnwnwnwnw ........AAAAAAAAIAIAIAIAIAIAAIAIAIAIAI iiiiiiiiissssss s s s s s ssss bebebebebebbbbbbbbebebebebebebeegigigigigigigigigigigigigigigigigiinnnnnnnnnnnnnnnniininininininininini ggggggggggggg tototototototooototototototototo dddddddddddddddddddemememememememememononononononononon tttttststststststststststs rararararararararaarateteteteeetetetetetetetetetete iiiiiiiiiitstststststststststststststs gggggggggggggrerererererererererereree tttatatatatatatatatatatatatatatat ppppppppppppppppppp tttotototototototototototottenenenenenenenene ttititititititititititialalalalalalalalalalalala aaaaaaaaaaaaaaassssssssssssss a a a a a a a a aaaaaaa ssososososossosososososososos urururururururururururcecececeeececececececececee oooooooooooooof f ff f f f f f ffffff f f hohohohohohohhhhhhohohohohohohopepepepepepepepepepepepepepepepeeppp ffffffffffffforororororororororrrororrrrrrr uuuuuuuuuuuus,s,s,s,s,s,s,s,s,s,s,s,s,s,s,s,ss llllllllllivivivivvvivivivivivivinininininininining g g gg gggggggg ininininininininininin ununununununununununu prprprprprprprprprprprppppppp dddededededededededededdededededediciciciciciccicicicicici tatatatatatatatatatatataablblblblblblbbbbblblblblblleeeee e e e e e e e eeee tttttitititititititititit mememememememememeessssssssssssss ofofofofofofofofofofofofofofofofofofof rrrrrrrrrrreeeeeleleleleleleleleleleleleneneneneneneene tltlttltltltltltltltlesesesesesesesesesessssssssssssssss hchchchchchchchchchchchchchchchchchanananananananana gegegegegegegeeeegegegegegee.......

Artificial intelligence (AI) has started to move beyond the realm of computer science and find genuine practical uses.What impact this will have on business and other parts of society is a matter of growing interest.A general-purpose AI developed by Hitachi is being put to use in a wide range of industries where it is successfully using past data to generate knowledge for a better future and to devise accurate ways to deal with the problem of unknowns.In doing so, it is giving a foretaste of new ways of confronting the unknown.AI is beginning to demonstrate its great potential as a source of hope for us, living in unpredictable times of relentless change.

　　　

84 Hitachi Review Vol. 65 (2016), No. 6

- 6 -

Approximate Solutions and True Solutions

Noriko Arai, Ph.D.Professor, National Institute of Informatics

Born in Tokyo. Hitotsubashi University Faculty of Law, graduated from University of Illinois, and completed her doctorate at the University of

Illinois Graduate School of Mathematics. Ph.D. (Science).

Dr. Arai specializes in mathematical logic (proof theory) and artificial intelligence.

Following positions as Hiroshima City University assistant, and visiting researcher at Fields Institute for Research in Mathematical Sciences and

the Princeton Institute for Advanced Study, she has been a professor at the National Institute of Informatics since 2006, and Director of the

Research Center for Community Knowledge at the same institution since 2008. She was awarded a 2010 Commendation for Science and

Technology by the Minister of Education, Culture, Sports, Science and Technology. Project director of the Todai Robot Project “Can a Robot Get

into the University of Tokyo?” since 2011.

Important works “Math is Language—Math Stories” (Tokyo Tosho Co., Ltd.), “How Computers Can Take Our Jobs” (Nikkei Publishing Inc.), “Can

a Robot Get into the University of Tokyo?” (East Press), etc.

Expert Insights

I don’t like machines very much. I don’t like riding in cars or on trains, and I chose a high school and

university that I could commute to by bicycle.

However, of all the various types of machines around today, I think I most dislike appliances with artifi cial

intelligence (AI) functions—‘talking appliances’ in particular. People are surprised to hear this coming from

the leader of an artifi cial intelligence project, the Todai Robot Project. Naturally there is no microwave in my

house. Microwave heating ruins tasty sake and makes mochi rice cakes lose their shape. It may be more

work, but sake is better when warmed in hot water, and mochi is better when cooked on a grill.

I have friends who say microwaving is handy. Most dishes can be cooked just by following the directions

of the appliance. Today’s microwaves have dozens of pre-programmed recipe settings, letting you cook things

like hamburgers and deep-fried shrimp just by doing as the appliance commands. Of course, microwaves

can’t actually deep-fry, so deep-fried foods are only simulacra of the genuine article, though reportedly very

tasty nonetheless.

However, herein lies the issue. When an approximate solution to a problem is provided, many people at

fi rst notice how different it is from the true solution, feeling, for example, that mochi is better cooked with a

grill. However, when they consider the labor required to achieve the true solution (cooking mochi with a grill),

people often just get lazy and accept the approximate solution (microwaving the mochi). The thinking goes as

follows: “I am clearly aware of the difference between the true solution and approximate solution, however on

this occasion I have simply made a rational choice by considering the costs versus the benefi ts. Naturally, I

will choose the true solution when I have to.”

Hitachi Review Vol. 65 (2016), No. 6 85

- 7 -

In his book, Laws of Media, the philosopher Marshall McLuhan makes the astute observation that this

way of thinking is a fl agrant misapprehension. Humans are shaped by the tools they produce, not the other

way around. By using these tools, we inevitably become a new entity that complements them. Socrates

also understood this. He refused to use the greatest invention in the history of mankind—writing. His choice

was lamented by his student Plato, who wrote down Socrates’ ideas. Plato’s choice may have been rational,

however we are unable to judge the value of what was lost by it since we are all the products of cultures from

after Plato’s time, so we have lost the ability to feel how different the world would be without writing. In the

future, we will likely be shaped by microwaves, by smartphones, by Watson*1, and by Siri*2.

So, when I am asked in interviews about whether artifi cial intelligence will become able to feel emotions

like humans do, I respond as follows: “When artifi cial intelligence becomes part of daily life, humans will

lose the ability to feel emotions before artifi cial intelligence gains them, so we will likely no longer be very

aware of the difference.” Journalists must not understand this answer, because it has so far never appeared in

any article.

*1 IBM and Watson are trademarks of International Business Machines Corporation, registered in many jurisdictions worldwide.

*2 Siri is a trademark of Apple Inc., registered in the U.S. and other countries.


Technotalk

- 8 -

The use of artifi cial intelligence (AI) is growing, with applications that range from big data analytics to marketing and self-driving vehicles. It is anticipated that new possibilities for business will open up through the combination of AI with large quantities of data acquired from real life, including human behavior and operating conditions from a variety of equipment. Anticipating such developments, Hitachi was among the fi rst to work on the development and deployment of general-purpose AI. Through this work, Hitachi is seeking to revolutionize business and society using AI solutions based on proprietary technology and concepts. Today, we invited Kazuto Ataka of Yahoo Japan Corporation, a leading evangelist for new AI, and engaged in a discussion with Kazuo Yano, who directs general-purpose AI research at Hitachi, concerning the shape of the future in which AI and people will co-exist.

A Number of Problems for AI

Yano: Currently, interest in AI is growing around the

world. Having had much to say about AI and data

analysis, drawing on a background of knowledge from

neuroscience, how do you view recent developments

in the fi eld of AI?

Ataka: As I see it, there are three problems. The

fi rst is the misdirected hype surrounding the subject

that, lacking an understanding of what AI actually is,

has engulfed it in fear-mongering and exaggeration.

Certain ways of thinking specifi c to Japanese people,

inspired by robot anime and a fondness for science

fi ction, may play a part in this.

The second is a lack of understanding of the

genuine changes brought about by the synergy of

AI and data. It is almost certain that many different

types of work based on the processing of information,

including the assessment and classifi cation of

information, analysis and prediction, and manual

work, will be automated at speeds tens of thousands

of times faster than human beings are capable of.

I wonder how many people genuinely appreciate the

incredible power of this.

The third problem is that, perhaps infl uenced by

these preconceptions and lack of understanding, and

despite Japan having many AI-related technologies,

I have a strong sense that we are lagging behind the

advanced economies of Europe and America, who

have a more realistic view of things and where the

pace of development is accelerating.

Yano: Having discussed AI in many different

places, I too, have become strongly aware of the

preconceptions you are talking about. The greatest

problem, which is not limited only to Japan, is the view

that humans and AI are in confl ict. This is the biggest

mistake.

Ataka: That’s right. And this is despite AI being for the

benefi t of humans.

Yano: The real confl ict is not with machines but with

other humans. It is a battle between the traditional

approach of only learning from one’s own experience

or from those around oneself, and the new approach

of utilizing computing power to learn systematically

from all available data. This is because it typically

makes no sense to compare the speed of a human to

that of a car, or the extent of your knowledge to that of

an Internet search engine.

Ataka: Because you can’t win comparisons like that.

Yano: A search engine is in truth an agglomeration of

AIs. In the same way, machines will always win out in

certain areas. Nevertheless, there are also things that

only we humans can do. I believe we need to increase

the number of people who view AI in terms of this

correct framework.

Ataka: I completely agree, and for this reason we

are engaged in public education initiatives. On the

other hand, when you consider the relationship with

robotics, there is also a danger that we will build

something that will be antagonistic to humans. What

are your thoughts on this?

Yano: I believe that ethics will become more important.

However, if you think about it, all tools are capable of

both good and bad uses, and creating a technology

that is absolutely incapable of being misused is an

impossible objective. If you consider AI as a tool, then

Kazuto Ataka, Ph.D. Chief Strategy Offi cer, Yahoo Japan Corporation

Kazuo Yano, Dr. Eng. Corporate Chief Scientist, Research & Development Group, Hitachi, Ltd.

Human-friendly AI that Learns from Life


- 9 -

Ataka: This extremely high capacity for learning is

not the only thing that the immune system has in

common with the central nervous system (CNS),

which I originally studied. The proteins expressed

on membranes are also similar. In fact, there are a

number of proteins that are specifi cally expressed in

the CNS and immune systems.

Yano: No doubt there is some signifi cance to that.

Along with the brain, I also see the immune system

and evolution as examples of knowledge-based

activity by living organisms. This makes me think that

AI research should pay more attention to mechanisms

from these two fi elds.

Ataka: That is a good idea. Although a life scientist

myself, I have just learned something new about living

organisms from a physicist, Dr. Yano (laughs).

Yano: I wouldn’t think of trying to teach you that.

Despite being a physicist, I have a fascination with

life and the study of it, and accordingly I have paid a

lot of attention to biology, especially the biology of the

human body, in my AI research. The question of how

we can get AI to understand human happiness and

use it to enhance that happiness is one of my main

research topics, and a key factor in assessing people’s

happiness is their bodily rhythms.

For the last 10 years we have been working on

research in which we analyze people’s activities

by using wearable nametag sensors to record

their movements in the form of three-dimensional

acceleration data. In doing so, we have identifi ed

characteristic patterns of bodily movement that

correlate strongly with people’s happiness. Looking at

fl uctuations in the duration and frequency of activity,

we fi nd that in groups with high happiness these

fl uctuations have a natural distribution, by which we

mean that there is diversity in people’s movements.

Through this analysis we have developed the ability

to numerically calculate a happiness index from

group activity data.

From various studies into variations among

living organisms, it can be seen that there is a

commonality in the rhythm of variation between

people and mice, and even with fl ies. Furthermore,

if mice or fl ies are genetically modifi ed to exhibit

characteristics of depression, the same disruptions

to rhythm can be seen in them as in people with

symptoms of depression. While we tend to think of

happiness as something that belongs to the fi eld of

psychology, I believe it is linked to more fundamental

aspects of biology.

Ataka: Behaviors (actions) you say? I believe our

emotions and thinking follow from our behaviors

(actions). From a neuroscience perspective, nerve

it is no exception. That said, there will be a need to

impose some form of restrictions.

Ataka: As with genetic engineering, there is a

need for measures that govern the technology.

The Boston-based Future of Life Institute, a research

support organization, is involved in activities such

as supporting research into AI that will contribute

to the future of humanity and looking into the risks

that AI poses. The Future of Humanity Institute at

the University of Oxford headed by philosopher

Nick Bostrom is studying the impact that AI will

have on humanity and how to control it. I have also

heard of similar institutions being established at The

Massachusetts Institute of Technology (MIT) and the

University of California, Berkeley. I believe we also

need to be doing something similar in Japan.

Yano: What is needed are places where people from

all walks of life can come together to debate the

subject, including those with no involvement in AI

research. More fundamentally, it may also be that we

need to be thinking about mechanisms to prevent the

misuse of technology or runaway research.

AI Should Mimic the Immune System

Ataka: While it is difficult to impose restrictions that do

not impede progress, there is a concern that, unless

debate gets underway and we start thinking about

guidelines soon, it will be too late. For example, a more

pressing problem than the misuse of robots may be

the use of AI in cyber-attacks.

Yano: But by the same token, can’t AI also be used

as a defense? In practice, consultations along these

lines are already in progress. Might it not be possible

to build AIs that act as an immune system, protecting

against unauthorized activities?

Ataka: Constantly retuning itself to provide protection.

That is quite a good fi t with how AI works.

Yano: So you agree that there is a great deal of

similarity between an AI and an immune system?

Ataka: Very much so. They both learn from a complex

environment.

Yano: The immune system deals with unknown

threats yet is based on a fi nite number of genes. It

works by partially accepting the unknown entity and

acquiring the means of manufacturing antibodies from

the invader itself so that it can be defeated the next

time it is encountered. It operates autonomously on a

completely different level than consciousness and the

brain and is eminently systematic. While the brain is

frequently used as a comparison in the discussion of

AI, it seems to me that we should be trying to mimic

the immune system.

88 Technotalk

- 10 -

about people as “having guts” or having a “gut feeling.”

In English, the meanings of the word “guts” include

grit and determination. In this way, it seems to me that

what an internal organ essential to our survival and a

decision-making system have in common are a gut-

like nervous system and a router.

Ataka: I agree. That’s because the nervous system

for the gut region is as well developed as the central

nervous system. In other words, the gut is also a brain.

Yano: The gut, too, is a brain.

Ataka: It seems that pursuing the brain of the gut is

one path open to AI research in Japan.

Yano: That’s right. That’s because we at least have a

capacity for “belly talk” (in Japanese idiom, the ability

to communicate through one’s attitude rather than

through words).

Ataka: That’s right. In the Ministry of Economy, Trade

and Industry (METI) council on which I sit, we are

currently debating the possibilities and challenges

of AI. A major topic in this discussion relates to what

Dr. Yutaka Matsuo of the University of Tokyo calls

“adult AIs” (information processing systems) and “child

AIs” (motor systems). This is the view that Japan

should leave adult AIs to major international players

and instead take advantage of our position as a world

leader in manufacturing to focus on child AIs that can

be used in construction, factory, and other workplaces.

As I see it, adult AIs correspond to the cerebral

cortex and child AIs correspond to the cerebellum.

If the ingenuity and manufacturing excellence of the

Japanese people are equated to the cerebellum,

then it is natural for us to direct our efforts toward the

cerebellum AI. The idea of a gut AI has never arisen.

Yano: Whatever it is that determines our gut feelings,

it is a decisive force. That people around the time of

the Meiji Restoration who didn’t know much English

could engage with people from America and Europe

on an equal footing and be respected was likely

because of their gutsiness and the fact that this is a

universal that goes beyond language. I believe that

AIs that make decisions on a larger scale rather than

competing on speed of detailed decision making may

be an option that plays to Japan’s strengths.

Ataka: The pitcher Hideo Nomo was playing major

league baseball when I was a graduate student in

the US, and I remember how a friend in my class

commented on how dignifi ed he appeared. Despite

not being able to speak English, he conveyed an

impression of having guts. Still, how do you go about

developing a gutsy AI? Yano: Perhaps it is something

we could work on together (laughs).

Ataka: It is certainly a very creative subject. It is

completely different from the usual concept of AI.

cells tend to die if they are not part of a network,

which is to say, if not connected to inputs and outputs.

Considering this, it may be that the body does not

exist for the benefi t of the nervous system, rather it is

the body that is central.

Adult AIs, Child AIs, and Gut AIs

Yano: That’s why we focus on acceleration. We want

to look at the outputs. Naturally, the ease of making all-

encompassing measurements is also a factor.

Ataka: If you look at the outputs, then the brain is not

all that important. This is a very bold idea.

Yano: The brain is like a router. Not that routing isn’t

important.

Ataka: “The brain is a router” – that makes a quotable

phrase.

Yano: What is important about routers is that different

tasks are performed through the same paths. We talk

Kazuto Ataka, Ph.D.

Chief Strategy Officer, Yahoo Japan CorporationJoined McKinsey & Company after completing the Master’s program at the University

of Tokyo in Biophysics & Biochemistry. After working there for four and half years,

he entered the Interdepartmental Neuroscience Program at Yale, where he earned

a Ph.D. in Spring 2001. After postdoctoral studies, he came back to Japan to re-join

McKinsey at the end of 2001. As a core member of its Marketing and Sales Practice

for the Asia-Pacifi c region, he was involved with brand rebuilding and product and

business development for a wide range of sectors. In September of 2008, he moved

to Yahoo. Following positions as Director of COO office and Head of Data, Research

and Strategy, he took up his current appointment in July of 2012. In addition to

resolving business strategy issues and promoting large-scale partnership projects,

he is in charge of the Marketing Insight and Intelligence Department, the Yahoo! Big

Data Report, and company-wide strategy including data utilization. Director of the

Japan DataScientist Society. Director of the Japanese Society of Applied Statistics.

Among his literary works is “Issue Driven – A Simple Essence of Intellectual Works”

(Eiji Press).


- 11 -

unknown world. Because we have no knowledge of

unknown problems, there is nothing to be gained

by repeating past practice. If you think about it

in terms of a hierarchical structure where, in the

past, a mechanism existed whereby a particular

phenomenon occurred and that mechanism in turn

was caused by a higher level mechanism, then you

realize that structure exists in the unknown, that a

way exists for fi nding a solution, and that each piece

of information contains a hundred meanings. I believe

that AIs capable of doing this in an extremely

systematic manner will prove to be partners that will

help people by building an as-yet-unknown future.

Ataka: But if AIs did this for us, would there be

anything left for people to do?

Yano: You can also view it as a means for better

utilizing the abilities we have. In relation to the

happiness index we talked about earlier, what we

found in a project at a call center was that the overall

happiness of the group was infl uenced by the level

of conversation during breaks, and that the order

Being Issue Driven

Yano: While deep learning, for example, has huge

potential, most real-world problems cannot be solved

by the weight of data and computing power alone. What

are needed are techniques that can deal with business

and other parts of society from a different angle.

Ataka: In this respect, while it may sound odd coming

from a director of The Japan DataScientist Society,

there are also aspects of the data science boom itself

that make me concerned. While data processing

techniques are clearly powerful when seeking to solve

problems in business, how to pose the problem is so

much more crucial, rather than how to apply specifi c

techniques or logic.

Yano: Which is the same idea you expressed in your

book, “Issue Driven – A Simple Essence of Intellectual

Works.” Most of the time when we fail it is because we

have misunderstood the issues.

Ataka: It is the same with me. The title of the book

was directed at myself (laughs). The majority of

business decisions do not require particularly

advanced data science. In this data-driven era, without

having a clear understanding of the issues, namely

what questions we are seeking to answer, we will

end up going in the wrong direction. In achieving this,

it is vital to get the basics right, meaning collecting

clean data and making appropriate comparisons in

accordance with the issues. I am concerned that

infl ated expectations for data science are spreading

through a failure to understand this point.

Yano: In that case, it is important to aim for AIs that

are capable of general-purpose application in a

variety of systems and that can respond proactively

and autonomously to all sorts of threats and other

changes, just like the immune system.

Ataka: I agree. Along with gutsy AI there is also

immune system AI. I will suggest at the next council

meeting that Japan should compete on the basis

of the cerebellum, the gut, and the immune system

(laughs). However, what is the current situation in

Japan? Are we ready to join the international battle for

AI development?

Yano: In the context of our earlier discussion of

issues, I believe we Japanese need to hone, not only

our skills for solving specifi c problems, but also those

for formulating good problems and concepts. Rather

than insights on their own, we need to strengthen our

ability to draw on specifi c experiences and formulate

concepts and problems, and I believe we can do so.

The formulation of good problems is in some

ways related to AI. We spoke about the immune

system, and I see AI as a tool for controlling an

Kazuo Yano, Dr. Eng.

Corporate Chief Scientist, Research & Development Group, Hitachi, Ltd.Joined Hitachi, Ltd. in 1984. In 1993, he achieved the world’s fi rst successful operation

of single-electron memory at room temperature. Since 2004, he has taken the

lead in the collection and utilization of big data. His papers have been cited 2500

times, and he has 350 patent applications. The wearable sensor he developed has

been described by the Harvard Business Review as a “historic wearable device.”

His literary work, “Invisible Hand of Data: The Rule for People, Organizations, and

Society Uncovered by Wearable Sensors” (Soshisha Publishing), was elected one of

BookVinegar’s 2014 10 Best Business Books. Dr. Yano has a doctorate in engineering,

he is an IEEE Fellow, a visiting professor at the Tokyo Institute of Technology, and

a member of the Ministry of Education, Culture, Sports, Science and Technology’s

Information Science and Technology Committee. He has been awarded many

international awards, including the 2007 MBE Erice Prize, and Best Paper at the 2012

International Conference on Social Informatics.

90 Technotalk

- 12 -

rate was 34% higher on days with above average

happiness compared to below-average days. The

ability of people to put their skills to use is infl uenced

by small changes in their surroundings.

One thing that was deeply interesting was that

there was no correlation at all between days with high

order rates and days with a high proportion of staff

with high order rate performance in the data from the

previous half year. You cannot build a strong team

from number four batters alone. The performance

of a team is not the sum of the individual skills of its

members. Rather, there are other factors that come

into play. One example is that there are people who,

despite their own results being poor, act to bring out

the best in those around them. Although we naturally

assume these subtle intra-group relationships must

exist, thanks to data and AI, this is the fi rst time we

have been able to visualize them. If we think only

about logic, all of these subtle effects tend to get

swept away.

Ataka: And yet the truth was present in the data.

Evolving AI for Exploring an Unknown World

Yano: Hitachi underwent a corporate reorganization

this April. Following the Research & Development

Division, which had already moved to a new structure

that focuses on collaborative creation with customers,

this involved a major transformation in which the

business units switched from the previous product-

based company structure to a customer-oriented

and service-based structure. Within this structure,

instead of selling AIs as such, our philosophy is to

supply services that utilize AIs to help overcome a

wide variety of customer challenges. Finally, please

tell me your views on this relationship between AIs

and services.

Ataka: In the case of interpersonal services, for

example, I see the greatest potential for data and

AI analysis coming at the stage of learning about

the background and latent needs of customers. In a

healthcare context, this would be when conducting

a diagnosis prior to writing a prescription. As for the

subsequent marketing and actual service delivery (or

prescription writing and treatment), because having

people do this work generates more value, I see AI

as serving a backup role, supporting people on the

frontline.

Elsewhere, data analysis can provide support

for work that in the past has been more of an art,

such as optimizing logistics or the physical layout of

equipment, etc. While these are back-end services,

they should enable improvements on the front end

through AI and people thinking together. While it is

difficult to simulate models that involve the interaction

of multiple systems, I see scope for AI in fi elds such

as business dynamics.

Yano: I see change as being the essence of services.

This includes both changes over time and changes

due to people or circumstances. In this sense, you

can describe AI as a tool for dramatically reducing the

cost of adapting to these changes. It is the generation

of wealth and elimination of inequalities in the world


- 13 -

through the resulting fundamental improvements in

productivity that are the ultimate goal of AI research.

Ataka: Nevertheless, the view persists that this is

something that only large corporations can achieve.

Yano: I don’t believe this myself. Whereas the options

for responding to change under past business logic

have been limited, data and AI open up a much larger

range of options. It can also be said that the process of

earning a profi t equates to fi nding a place for oneself

in the network of economic transactions that is of

benefi t to everyone. It is my hope that the use of AI to

look at things in ways that are different from those of

the past will enable a greater number of companies to

fi nd such a place for themselves.

Ataka: Finding their own niche, in other words.

Yano: And isn’t that just another way of saying

evolution?

Ataka: That’s right. In ecological models, the opening

up of a new niche leads to the emergence of new

species.

Yano: Evolutionary AI supports this process.

Ataka: Yes, I see the connection. I hope we can make

ongoing progress on research into evolutionary AI that

will open up unknown worlds.

Yano: Our aim is to create human-friendly AIs that

work with us to create a happy future. Thank you for

your time today.


- 14 -

Featured Articles

AI for Taking on the Challenges of an Unpredictable Era


RISING INTEREST IN AI

ARTIFICIAL intelligence (AI) is a topical subject. Large investments in the technology, which is a major factor in international competitiveness, have been announced by governments and other organizations.

Nobody as yet has an overall grasp of what impact AI will have on business and other parts of society. However, new developments that give a glimpse of the future have already begun to appear. These can be found in marketing, academic papers, investment decisions, and Internet chatter. And they already include examples of how things will play out in the future. The future is already happening.

However, very few people have the opportunity to deal with these in their entirety. Business people find it difficult to understand technology and its significance while, for AI technologists, the world of business is far away.

The author’s own life reached a turning point some 13 years ago when Hitachi’s exit from the semiconductor industry meant that he was forced to abandon the semiconductor research he had been engaged in for the previous 20 years. While the company’s decision was very unfortunate for him personally, he regrouped and made a fresh start along with his colleagues. This was a turning point because it enabled him to embark on research into what is now called big data and the Internet of Things (IoT), along with AI, before these became popular topics. It wasn’t that he had any particular vision, it was more a matter of someone who had nowhere else to turn finding unexpected strength. Looking back, there is nothing for which he is more grateful than this decision by Hitachi to shut down its semiconductor business.

Thanks to these changes, the author now gives more than 500 lectures and other presentations on AI each year, providing opportunities to meet many different people and to discuss how AI relates to companies, business, and people. These range from lectures given to audiences of 1,000 or more to the board meetings of listed companies. The people he

meets come from a wide range of industries, including many executives from the manufacturing, finance, retail, logistics, and public sectors.

Through these discussions he has been able to identify developments that, while still in their early stages, are already up and running and will have a significant impact in the future. This article describes the developments he has found.

MISUNDERSTANDINGS ABOUT AI

Accompanying the recent rapid growth of interest in AI there have been increasing numbers of editorials and other commentaries coming out that the author finds disquieting. The following are statements that are commonly heard in relation to AI.

“Machines can now beat humans, even at the game of go.”“This will lead to competition between people and machines (AI).”

Both of these statements are misleading.The science journal, Nature, carried an article about

how the AlphaGo go-playing software developed in the UK defeated the European champion(1). While computers have already beaten professional players of both chess and shogi, because the search space of go is orders of magnitude larger it was believed that victory would not come so easily. Nevertheless, computers have become triumphant in a much shorter time than expected. Because the software uses deep learning(a), a topical AI technique, it has been reported that AI has now surpassed human intelligence.

By contrast, the author sees this in terms of a battle between humans.

(a) Deep Learning A machine learning technique that incorporates mechanisms from the

neural circuits of the brain (deep neural networks). Like the brain, deep learning is designed to increase the weighting of circuits that produce correct answers and it can make judgments on unknown patterns, especially in images, by having the computer identify features on its own from input learning data.


- 15 -

On one side are those who adopt the traditional approach of improving their abilities through experience and learning. In other words, people who choose to compete using their own physical and mental strengths.

On the other side, there are those who choose to learn systematically from past records of games, using computers for this purpose, and then having these computers play against each other tens of millions of times to provide more data for learning. These people take a systematic approach to thinking about how best to take advantage of large amounts of past data and the overwhelming data processing and memory capabilities of computers, and apply their physical and mental strengths to this end.

In both cases, it is a human choice and the competition is one between humans.

As a result, those who have taken the latter approach have predominated. That is, success has been achieved by striving systematically to build computer-based techniques for tackling the problem of unknowns.

The reason for this is simple. It is because computer performance has improved and more data has become available to use for learning. This does not apply only to the game of go. The same thing is also happening in business.

An important factor for business in all of this is that the development team for the go program did not include anyone with professional-level skills in the game. The development of conventional business and management systems has required expertise in the relevant fields. By contrast, the systematic learning used by the computer described above did not require the people involved to have any special expertise. It is anticipated that, as greater use is made of systematic learning, the relative value of practical application-specific knowledge will diminish.

Instead, what is important is data. The value of business and other real-world data is growing rapidly. What is also becoming important is the ability to use computers to learn systematically from data.

To view this as a competition between humans and machines is to misunderstand what is happening. Despite the fact that new ways of solving problems through AI are driving rapid changes in what is required of people and in how they go about living their lives, there is a risk that they will adopt misguided ideas and actions by failing to recognize this.

Nobody would be surprised to hear that the track athlete, Usain Bolt, lost in a race against a motor vehicle. Likewise, the author has never heard anyone

claim that their own knowledge is less than the results of a web search engine (which itself is a form of specialist AI). People are simply happy to make use of vehicles and search engines. And, by using the technology themselves, people will acquire an understanding and rid themselves of this unusual way of looking at the situation as a competition between humans and machines.

NECESSITY OF AI

Use of AI as a new methodology is growing rapidly in business. This is because of the significant positive impact it has on productivity.

20th Century and StandardizationPeter Drucker predicted that “The most important, and indeed the truly unique, contribution of management in the 20th century was the fifty-fold increase in the productivity of the manual worker in manufacturing. The most important contribution management needs to make in the 21st century is similarly to increase the productivity of knowledge work and the knowledge worker(2).” In other words, Drucker sees a fundamental difference in the nature of work in the 20th and 21st centuries.

The 20th century was one of dramatic improvement in the productivity of factory work. The driving force behind this was the scientific management theory of American engineer and management theorist Frederick Winslow Taylor. Taylor conducted rigorous studies of shoveling work at a steel mill. He broke the work down into separate processes that he then looked at individually to identify activities that were unnecessary or that could be done more quickly. Based on these studies the necessary processes were then standardized. This made it possible for work that had been believed to be possible only by experienced staff, to be done instead by inexperienced people while still maintaining a level of quality.

Taylor’s scientific management theory was adopted across a wide range of activities during the 20th century. This led to work associated with diverse tasks and services being broken down into processes and standardized to eliminate waste.

In the latter half of the 20th century, computers were introduced as a way to achieve this with even greater rigor. Once a computer program was written, it could be used to process and output large amounts of data. Initially used for accounting, computer applications have been expanded into all areas of corporate activity

94 AI for Taking on the Challenges of an Unpredictable Era

- 16 -

to monitor and automate such business processes as order entry, procurement, production, inventory, dispatch, and human resources. In accordance with Taylor’s philosophy, business activities were broken down into individual processes and standardized, with computers being used to record and manage the status and progress of each standardized process. This achieved a fifty-fold increase in productivity and created the modern economies of developed nations.

Now, however, with services and other knowledge work accounting for more than 70% of all work in developed nations, Taylor’s methods are no longer sufficient on their own. This is because of the extent to which changes have occurred in services and other knowledge work and the environment in which they are performed. Rigid and uniform practices are frequently unable to cope with people’s preferences, greater individual diversity, and the characteristics of specific locations or regions, not to mention fluctuations in things like demand and prices. Rigid business processes based on the Taylor model are a poor match with reality, and even defining business processes is difficult. Similarly, the definition and measurement of business productivity are also frequently difficult.

Benefi ts of Introducing ComputersIn parallel, this also means there is a limit to the productivity benefits that can be achieved by introducing computers as described above. Conventional computers do not learn and grow when circumstances change. For

this reason, they are referred to as hard-coded systems, because the programs have to be written explicitly.

In accordance with Taylor’s philosophy, people who improve their capabilities by sharing best practices are called Human 2.0, which is to say they are characterized as standardized workers. By contrast, people whose capabilities have been specialized through the division of labor are called Human 1.0, which is to say they are characterized as specialized workers (see Fig. 1).

Current corporate information systems have been developed to support the work of this second generation (Human 2.0). Unfortunately, this approach is reaching its limits in terms of cost-benefit. This is the background to the emergence of the third generation of machines and information systems described in this article.

The productivity of nurses or department store sales staff, for example, cannot be improved using manuals alone. In addition to their core role of caring for patients, nurses also need to produce documents and to consult and coordinate with other people. Likewise, in addition to their core role of recommending products to customers and encouraging their interest, sales staff also need to produce documents, check inventory, and keep track of deliveries. In a diverse and ever-changing environment, it is not possible for manuals to document things like how to prioritize this work and allocate time, meaning that the nurses and sales staff must make these decisions for themselves.

Human 1.0

Specialized Worker

First generation

Machinesthat convert energy

from 1769

InputPrerequisites

170 years

Prerequisites

This Work

70 yearsOutput

Fuel (coal)

Mechanical→ Electric power

Problem/objective

Decision-making/optimization

Program (processing procedure)

Data

Adam Smith(1723-1790)

Second generation

Machinesthat calculate

from 1940

Frederick W. Taylor(1856-1915)

Third generation

Machinesthat learn

from 2010

Human 2.0

Standardized WorkerHuman 3.0

Amplified Worker

This Work

Advances in human

capabilities

Advances in machine capabilities

Fig. 1—Improving Productivity through Advances in Human and Machine Capabilities.People who have built tools and have been specialized through the division of labor (first generation) amplify their capabilities by disseminating and learning from the know-how of experts (second generation), and by autonomously learning from real-world situations that transcend time and space (third generation).


- 17 -

What is important for services and other knowledge

work is to set work objectives and other goals. In this

article, these are referred to as outcomes. Given a clear

objective, it is possible to make flexible and accurate

decisions under widely varying circumstances. Past

information systems have not been able to satisfy this

need for flexibility. Instead, they have tended to result

in more standardized practices.

New Image of HumansWhat changes all of this is the advent of new methods

based on the use of AI. Modern information systems

are steadily building up huge quantities of data. The

changes taking place in businesses and other parts

of society are reflected in this data. By utilizing the

data, it is becoming possible for computers to modify

the logic by which they operate in accordance with

changing circumstances. In other words, AIs have the

ability to learn from data and make decisions.

Further underpinning this progress are developments

in data collection methods. Ongoing advances in

technology include sensors, wearable devices, robots,

and drones. Combined with AIs that can learn from

this data autonomously, this can create systems (AI-

based systems) capable of adapting flexibly to change.

These new AI-based systems have the potential

to provide powerful augmentation and amplification

of human learning abilities, and can be expected to

deliver productivity improvements. This represents

the third generation, Human 3.0. A feature of this

generation is the amplification of human capabilities.

In other words, an amplified worker (see Fig. 1).

Whereas enterprise resource planning (ERP) and

other similar computer systems were used for

standardization in Human 2.0, AI-based systems that

learn will assist people in the third generation, Human

3.0, with ongoing learning from quantities of data that

are too large for humans to consider on their own.

This will enable decision-making with an accuracy

that would be impossible when relying only on human

experience. It will also allow flexibility in decision-

making in ways that are not able to be documented

in conventional procedure manuals such that systems

will be able to adapt if business conditions change

(such as distribution channels or supply and demand

conditions). This contrasts with the tendency for

people to persist with practices that have worked in

the past, even when circumstances change.

Accepted wisdom in the past has been that

business process standardization and procedure

documentation represent best practices for improving

business efficiency, and customers and staff have

adapted to work with predefined procedure manuals

and machines. In practice, however, the policy of

standardizing business processes and implementing

them on a computer has not been a success in many

services that need to deal with change and a wide range

of other conditions. In fact, the concern has been that

they pose an obstacle. This in turn has impeded the

growth of the traditional information system business.

In the third generation of work, it is the computers

and other machines that adapt to humans rather than

humans adapting to machines and processes. They

support people who make autonomous decisions amid

changing circumstances and who are accountable for

the results.

AI AS A WAY TO SURVIVE AN UNPREDICTABLE FUTURE

The arrival of AI represents more than just the

provision of a convenient tool. Rather, it changes

how we go about solving problems and other aspects

of our life. While the future is full of endless new

possibilities, it also harbors threats to our existence.

In talking about the impossibility of predicting the

future, Peter Drucker offered the following insight.

We know only two things about the future:

It cannot be known.

It will be different from what exists now and from

what we now expect.

(...)

The purpose of the work on making the future is

not to decide what should be done tomorrow, but

what should be done today to have a tomorrow.(3)

In other words, we are only able to act today. We

have no direct means of changing the past or the future.

We can only influence the future by how we act today.

However, the author believes there are two different

ways of thinking about the present (the point where

past and future meet).

The first is to see the present as resting on the

knowledge and events of the past. Under this view,

what is important is to study established knowledge

(academic, scientific, and practical knowledge) and

put it to use today. When confronted with things not

covered by past knowledge, the only option is to deal

with it as it presents itself.

The second is to view the present as the vanguard of

an unpredictable future. In response, even if the future


- 18 -

Specifically, xian, lin, and jian, the Chinese

characters used in the above names, are three

hexagrams from the full set of 64, something that

was well known among learned people of that time.

The xian hexagram refers to conjoining with space

and other aspects of the world in ways that go beyond

words. The I Ching classifies changes using hexagrams

that can be specified by a six-bit binary number, with

changes being represented by a combination of two

states, yin and yang. If a binary representation is

adopted in which 0 is yin and 1 is yang, the code for

xian is 001 110. The term lin (110 000) refers to acting

forthrightly to unhesitatingly grasp opportunities that

arise. Accordingly, Kanrin (the Japanese translation of

xian – lin) can be interpreted as indicting a situation

in which one should be conjoined with the world in

ways that go beyond words, and by doing so, be ready

to act forthrightly to seize opportunities that present

themselves. In other words, by using the I Ching’s

language of hexagrams to name his vessel, the name

expressed the attitude of facing up to an era of new

changes in the transition from the Edo to the Meiji

Period.

Similarly, the jian hexagram (001 101) referenced

by Mutsu means to progress one step at a time despite

adversity by asking for help from others. The Kenken

in the title of Mutsu’s book is the Japanese translation

of jian-jian, with the repetition serving to emphasize

this meaning. That is, it means to progress one step

at a time despite adversity by asking for help from

others, such that even if this leads to more adversity

one can continue to make steady progress by asking

for help from others. In this way, a large number of

changes (64×64 = 4,096) can be expressed by pairs

of hexagrams.

For learned people of the Edo and Meiji periods,

the pursuit of learning was all about becoming the

sort of person who adopts the correct attitude to the

unknown. The I Ching was the canonical and foremost

text for this purpose.

Furthermore, this methodology of the I Ching,

namely the use of data as a basis for selecting from a

set of systematically predefined options, is the same

as that used by modern AI. The author discusses this

in more detail below.

It is completely different from the modern idea of

learning as being an understanding gained from sources

such as books or schools. The view that learning was

about the acquisition of pre-existing information

spread quickly, starting from around the generation

that came after those educated during the Edo Period

is impossible to predict, it should still be possible to

establish ways of dealing systematically with a diverse

range of changing circumstances. This view of the

present is about striving to achieve this.

The 20th century was a time in which scientific

discoveries were made in a wide range of fields,

with widespread technological applications. Because

knowledge was being created so rapidly that learning

found it hard to keep up, it was a time in which the

former approach to the present prevailed. It is an

approach that is deeply engrained within people.

AI is a new way of doing things that represents a

shift toward the latter approach to the present and also

a turning point that combines both.

Which of the two approaches has been emphasized

has changed over time. Historically, there have been

times in the past when dealing systematically with a

future that is impossible to predict was seen as more

important than it is now.

The Edo Period was one such time. Kaishuu Katsu

and Mutsu Munemitsu provide examples. Both of them

changed history in the transition from the Tokugawa

shogunate to the Meiji Era. They also had something

else in common, the ship in which Kaishuu Katsu

visited America was the Kanrin-maru. The main work

of Mutsu Munemitsu was his Kenkenroku diplomatic

memoirs. Each of these people chose the respective

names from the hexagrams described in the oldest of

oriental classics, the I Ching.

Written more than 2000 years ago, the I Ching

states that there are ways of dealing systematically

with the unknown based on the identification of

the smallest of signs. The scope of the I Ching’s

application extended from matters of state to personal

decisions. The other English name given to the I Ching

is The Book of Changes. It systematically classifies

unknown changes into 64 hexagrams and explains

how they are to be interpreted. Given its place as

the foremost of the Four Books and Five Classics of

Confucianism(b), the I Ching clearly placed ways of

dealing with the unknown as central to learning. Over

and above knowledge of practices and information

that have already been established, it placed an

emphasis on ways of dealing systematically with these

unpredictable situations.

(b) Four Books and Five Classics of Confucianism

Nine books that are recognized for their particular importance in the

teachings of Confucianism. The four books are the Analects, the Great Learning, the Doctrine of the Mean, and the Mencius. The five classics

are the I Ching, the Book of Documents, the Classic of Poetry, the

Book of Rites, and the Spring and Autumn Annals. Confucianism is

the general term for the ideas and beliefs attributed to Confucius.


- 19 -

These new techniques require broad applicability

(i.e. they need to be general-purpose) in order to be

able to deal with unpredictable situations. Put another

way, the progress of AI depends on the extent to which

it is capable of broad application.

PROGRESSIVE DEVELOPMENT OF AI

Along with the following five requirements associated

with responding to unanticipated changes, how broadly

AI can be applied depends on the extent to which these

can be generalized.

The five requirements for responding systematically

to unanticipated situations are as follows (see Table 1).

F0: Outcomes: Determining the objective to be

achieved

F1: Scope: Determining the scope that needs to be

considered

F2: Options: Devising a list of options (potential

actions)

F3: Decision criteria: Devising an evaluation

function that specifies the criteria for selecting

which action to take

F4: Decision and optimization: Using the evaluation

function as a basis for selecting which option

(action) to take

F1, F2, F3, and F4 are then revised based on the

outcomes of the action taken.

There is no great difficulty in developing a

program for a particular problem that satisfies these

five requirements. What is difficult is achieving the

broad applicability needed to use the program on

unpredictable problems in which the circumstances

are not known in advance.

While achieving broad applicability is more

difficult the higher the requirement is in the table

described above and continuing to the era of catch-up

and overtake in the Meiji and post-war periods.

Now, however, new AI methodologies are placing

a fresh emphasis on responding systematically to

unpredictable changes over and above the utilization

of existing knowledge.

Furthermore, as the essence of management, as

noted by Peter Drucker, lies in adopting the right

approach to an unpredictable future, AI will transform

corporate management.

NEED FOR AI TO HAVE BROAD APPLICATIONS

In practice, how is it possible to respond systematically

to unpredictable situations? And how can the

capabilities of computers help with this?

First, there is a clear need to use general-purpose

methods. Countermeasures built on guesswork will be

of no use in unanticipated situations.

The practice adopted in the past has been

categorization. The idea was that, by grouping situations

into categories and devising countermeasures to each

one in advance, the correct response could be delivered

when something happened. While this works well

for simple problems, it fails when the situation is

complex. The accurate categorization of situations

requires a large number of categories. Furthermore,

because it is in principle impossible to anticipate all

possible situations, there is an ongoing need to update

the categories and countermeasures. In many cases,

this takes too much work to be practical.

By contrast, rather than relying on predefined

categories, the new approach using AI involves loading

large amounts of data from past and present and taking

advantage of the overwhelming data processing

capabilities of computers to infer the appropriate

response from past examples.

Problem What to decide

F0: Outcomes What is the objective to be achieved? Outcomes (KPIs)

F1: ScopeWhat is the scope that needs to be

considered?Input data (scope)

F2: Options What potential actions are available? Options

F3: Decision criteriaWhat are the criteria for selecting

which action to take?Evaluation function

F4: Decision and optimizationWhich action should be taken

(decision)Actions

F1 to F4 are revised based on the outcomes of the action taken.

TABLE 1. Five Requirements for Responding Systematically to Unpredictable Situations

There is a need for broad applicability to unpredictable problems in which the situation cannot be foreseen.

KPI: key performance indicator


- 20 -

is multiple regression(d), a technique for determining

the parameters of an approximate model (gradient and

intercept) by fitting it to actual data (scatter diagram)

using the least squares method. Other forms of machine

learning include more complex forms of multiple

regression, such as support vector machines (SVMs),

decision trees(e), deep learning(8), and collaborative

filtering(f). In essence, however, these are just more

complex models that achieve greater accuracy

by increasing the number of multiple regression

parameters. These machine learning algorithms are

already well known and are available through open

source libraries such as the R language.

These level 1 systems are special-purpose AIs

that are necessarily designed for a particular purpose

only. Well-known examples include the generation

of customer recommendations in the retail sector

and systems for performing facial recognition from

photographs. A system for generating quiz answers(9)

and a technique for identifying a person’s face in web

(F1), once achieved, the higher the requirement is in

the table, the greater the possibility is of applying the

AI program in a wider range of situations.

An outcome is an indicator of success (how good

or bad a result is). In business these are also called

key performance indicators (KPIs). Choosing these

indicators is an important human decision. It is

something that in principle must be decided by people.

There has been considerable discussion and

research into artificial general intelligence (AGI) in

recent times. While the definition of AGI remains

unclear, it is frequently used to refer to artificial

intelligence that has a similar broad scope to human

intelligence. The idea that the achievement of AGI

will bring about a singularity has been widely

debated(4) – (7). In the sense used above, the pursuit of

broad applicability in itself is the right way forward

for AI. However, the author’s personal view is that

arbitrary distinctions about whether or not AGI has

been achieved are undesirable.

In other words, people should avoid making

arbitrary distinctions and aim instead to progressively

expand the applicability of AI. The five requirements

above are useful for this. The author has devised his

own classification for the progress of AI comprising

four levels based on its scope of application (see

Table 2).

Level 0 indicates traditional mechanistic systems

that do not include learning from data. These are not

AI. Level 0 programs are implemented as fixed logic

and are written by hand. Most existing corporate

information systems and infrastructure systems are

at level 0.

Level 1 indicates systems that autonomously

modify parameters based on data to achieve an

objective, and in which these parameters have been

specified by people. Most systems that use machine

learning(c) are at level 1. The simplest example of this

Level Category Features (outcomes are given) Examples*

4

General-purpose

AI

ScopeDecide by learning from data to determine scope and options

Corporate strategy, urban design, wellbeing policies

3 OptionsDecide by learning from data in a given scope to determine options

Generation of programs for inventing drugs or materials

2 JudgmentDecide by learning from data for a given scope and options

Optimization of factories, warehouses, or sales

1 Special-purpose AIUpdate specified parameters based on learning from data

Recommendations, answering questions

0 Non-AI Fixed logic specified by hand Existing business systems

TABLE 2. AI Levels

AI is classified into levels 1 to 4 based on its scope of application. Level 0 applies to existing mechanistic systems that are not AI.

AI: artificial intelligence * Examples for levels 3 and 4 are the author’s predictions

(c) Machine Learning

An AI technique that performs recursive learning from past data and

uses this to identify meaningful patterns. Predictions of the future can

be made by applying the results to new data.

(d) Multiple regression

An analysis technique for predicting a particular variable (objective

variable) using multiple variables (explanatory variables). In the case

where the value of purchases by a customer at a store is the objective

variable and the customer’s annual income, age, gender, and family

structure are the explanatory variables, to use annual income alone to

predict the value of purchases would be single regression analysis,

whereas using the customer’s annual income, age, gender, and family

structure to make the prediction is multiple regression analysis.

(e) Decision tree

A technique for assisting with decision making by using a model with a

tree structure to analyze the factors involved in achieving an objective.

In the machine learning field, decision trees are used as predictive

models in which the objective variable and explanatory variables are

represented by a tree structure. The technique is commonly used in

data mining.

(f) Collaborative filtering

A method for inferring the preferences of a particular user from their

history of activity that hypothesizes that user preferences will be similar

to those with a similar past history of activity and creates a database

of the relationships between the history of activity and preferences of

a large number of users.


- 21 -

into a system that learns and grows by specifying the

outcomes and the inputs and outputs.

The daily operation of a warehouse was improved

by connecting a general-purpose AI to its warehouse

management system (WMS). The outcome in this case

was the minimization of total working time. To achieve

this, H received the latest data from the WMS each

night and used it to determine an equation (evaluation

function) expressing how to minimize the total

working time. When the dispatch orders (numbering

in the thousands) were ready in the morning, H then

performed automatic scheduling to prioritize which

orders to execute first and output the results as a

picking list. While the warehouse staff would then

perform the work in accordance with these instructions,

in practice they would make various trial and error

initiatives during their work. The consequence of this

was that work time would decrease on some days and

increase on others. As this in turn became input data

for the AI, the schedule produced for the next day

would take it into account. In this way, integrating a

general-purpose AI into an existing system succeeded

in creating a learning arrangement whereby the AI

and staff cooperated on a daily basis. The warehouse

achieved an 8% improvement in productivity.

Other projects achieved a 15% increase in sales

at a retail store and a 27% increase in orders at a call

center. As described in this issue of Hitachi Review,

the technology is being deployed in applications such

as finance, railways, factories, and water treatment

plants, with a total of 24 projects having commenced,

covering seven different industries (see Fig. 3).

images(10) have been reported in the media, and both

of these are special-purpose level 1 AIs.

In the case of level 1 systems, however, the choice

of data (feature values) to input to the algorithm is

determined based on hypotheses made by humans,

requiring the development of a program, and problem-

specific program development is also required in

association with the algorithm. Accordingly, each

problem to which the system is applied involves

accompanying analysis and development costs.

Furthermore, because the systems lack general-

purpose capabilities (are unable to go beyond

hypotheses that involve feature values anticipated by

people), they are unlikely to produce results that go

beyond human ideas.

By contrast, level 2 AIs have the broad applicability

to overcome these limitations. Level 2 systems

generate decision criteria automatically from large

amounts of data and select the best option from among

those offered by people. In this case, the scope (range

of data to be considered) and outcomes are given by

people (see Fig. 2).

A major difference from level 1 is that both the form

of the evaluation function used for decision making

and its parameters are determined automatically from

the data. This means that, in level 2 systems, people do

not need to predefine the hypotheses about the subject

matter, and therefore they can generate solutions that

people had never considered.

Another difference is that there is clear

independence between level 2 AI and non-AI systems

(IT and other equipment) (see Fig. 2).

It is the broader applicability of level 2 AIs that

makes this independence possible. This leads to

significant savings on the cost of implementation.

Because level 1 learning algorithms are permanently

embedded in the system, there is no clear distinction

about where the AI starts and ends. Accordingly, there

is a large cost associated with program development

and maintenance for each problem, and only in rare

instances do the benefits (outcomes) outweigh the costs.

The Hitachi AI Technology/H (hereafter referred

to as H) announced by Hitachi in 2015 is a general-

purpose level-2 AI. As far as the author is aware,

it is the first such level-2 AI to enter practical use.

H automatically generates more than a million

hypotheses, identifies which factors are important,

and determines in a quantitative manner the conditions

under which better outcomes will be achieved. This

technique is called leap learning. A general-purpose AI

can be added on to an existing system to transform it

Monitoring

Control

Systems AI

Produce results while learning from data and growing in accordance with the situation

• Characteristic 1: Outcomes and inputs and outputs are specified by people

• Characteristic 2: No need to specify domain- or problem-specific logic

• Characteristic 3: Can operate as an add-on to existing systems

Existing systems

IT andequipment

Hitachi AITechnology/H

Fig. 2—Features of General-purpose AI (Level 2 and Higher).People specify the outcomes and the inputs and outputs (based on factors such as the problem scope) and the AI operates based on the data to find or control the conditions that improve the outcomes. There is no need to specify problem-specific logic. In general-purpose AI, the AI and non-AI systems are separate.

IT: information technology


- 22 -

growth strategies even amid unpredictable changes

in circumstances resulting from globalization, and to

generate the actions needed to guide people toward

their own wellbeing.

IMPACT OF GENERAL-PURPOSE AI

For the purposes of this article, AIs at level 2 or higher

are referred to as general-purpose AIs and those at

level 1 as special-purpose AIs. Level 2 AIs can already

be used on a wider range of problems than could be

covered by any individual person. The quantity of

data that can be loaded and knowledge that can be

extracted from it is much greater than can be processed

by one person. However, as noted above, this should

be thought of not as machines outstripping humans,

but as a measure of how far human capabilities have

been enhanced.

To demonstrate the power of these general-purpose

AIs, Hitachi used LEGO*1 blocks to build a robot

capable of swinging on a playground-style swing and

connected it to the H general-purpose AI. The outcome

(objective) was specified as maximizing the amplitude

of the swing. The system was set up so that movement

data was input into the AI via the controller and so that

the AI could control the movement of the robot’s knees.

The robot was successfully able to use the swing after

only about five minutes of system operation, without

the use of any predefined knowledge.

In this way, an existing system was transformed

into one capable of learning, applying what it has

learned, and growing simply by integrating it with a

general-purpose AI and providing the desired outcome

and data with the potential to influence that outcome.

Such systems based on general-purpose AI can

operate 24 hours a day, 365 days a year. Nevertheless,

The important point is that the same general-

purpose AI software (H) is being used without its

being customized for the specific industry or problem.

This has resulted in a sudden expansion in the scope

of application of AI.

While the system is a level-2 AI with a wider range

of application, it is still only capable of choosing the

best alternative from a range of options provided

by people. In practice, this is enough to implement

fairly complex control functions. In the warehouse,

for example, the AI prioritizes which of the thousands

of dispatch orders to perform first. The total number

of combinations for executing 1,000 dispatch orders

is nearly infinite (approximately 1,000 factorial). This

can be determined accurately.

However, producing more creative answers requires

the even greater degree of generality provided by a

level 3 AI. For level 3, the AI uses data to determine

the options for what action to perform next. An AI

with this capability would, for example, be able to

produce new drugs or other materials or write software

automatically based only on examples. What makes this

intrinsically more difficult than level 2 is the need to

determine the procedures or sequences for generating

these. Here, the term “procedure” means a complex

combination of things like conditional branches and

loops based on intermediate results, elements that are

not present in level 2. The inclusion of these causes

a dramatic increase in the size of the search space,

making it difficult to learn from example data.

Up to level 3 it is people who determine the scope

and decide which data to input. In level 4, the AI also

determines what data to input (the scope). With this

capability, the ability of AIs to choose their own inputs

enables a higher degree of generality to be to achieved

by combining AIs in complex ways.

If level 4 AIs can be achieved, it should be

possible to devise sustainable national and corporate

Retail

Identify optimal conditions for work and supervision

8% improvement in productivity

Identify optimal uses of staff time

15% increase in sales per customer

Identify efficient operating practices

Estimated 3.6% reduction in operating costs

Identify how to improve productivity by talking to staff

27% improvement in order rate

Logistics Seawaterdesalination plant Call center Fig. 3—Applications of H

General-purpose AI.H has been used in 24 projects across seven industries, including finance, retail, logistics, plant, transportation, and manufacturing, where it has generated actions and improvements in a general-purpose manner using the same AI software.

*1: LEGO is a trademark of the LEGO Group.


- 23 -

Meanwhile, this general-purpose technology(11)

(the term used by US economist Erik Brynjolfsson)

will come to be more broadly applied after initially

proving itself in specialist applications. This is a

very important turning point because this broader

applicability will greatly expand potential applications

while also driving down costs.

Mobile phones provide an example of this. The

first mobile phones were vehicle-mounted and enabled

people to call each other whenever they wanted.

Subsequently, the widespread adoption of smartphones

transformed the mobile phone into a truly general-

purpose device, thereby greatly increasing its value.

The size of the resulting market was much larger than

it had been prior to this wider scope of use.

Likewise, given that an evolution from special-

purpose to general-purpose AI can be anticipated,

Hitachi chose to start development of general-purpose

AI earlier than the rest of the world. This work is now

beginning to bear fruit (see Fig. 4).

AI MEANS SEARCHING A LARGE INFORMATION SPACE

Essence of AI TechnologyWhat is the nature of the technology that makes this

AI possible? The answer has already been given in

Computing Machinery and Intelligence(12), published in

1950 by Alan Turing, the mathematician who provided

the theoretical underpinnings for computers. Turing

they are able to learn quickly. It was astonishing that

it took only five minutes to learn how to use a swing.

This also demonstrates its ability to adapt to changes

in supply and demand, prices, people, places, or time.

Unlike people, it can continue doggedly to adapt without

requiring instructions. Moreover, the H general-purpose

AI also indicates the reasons for its decisions.

Past special-purpose (level 1) AIs required

new programs to be written for each problem. The

provision of product recommendations at a retail

store, for example, required dedicated software to be

developed. With a general-purpose AI, by contrast,

product recommendations and optimal product

ordering or product ranges can be achieved simply by

changing the configuration of the H software, without

new program development.

The emergence of general-purpose AI is expected

to have as great an impact as the initial development

of computers some 80 years ago.

The current interest in AI began about two years

ago. Nevertheless, behind the scenes, AIs have been

playing a part in people’s lives for the last 15 years or

so. When someone does a web search in response to

a product recommendation from an e-commerce site

they are using AI technology without their knowing

it. This is because AI technology is used behind the

scenes in both cases. However, these are both special-

purpose AIs designed specifically for the application.

The recommendation engine cannot be used for web

searches and vice versa.

Gen

eral

-pur

pose

App

licat

ion-

spec

ific

1960 1970 1980 1990 2000 2010

Quiz answers

Chess

Recommendations

ICOT established

Web search

Collaborative filtering

Deep learningBayesian inference

MC/particle filtering

Winograd (1987)

Dartmouth Conference

Expert systems

Fifth-generation computers

Pattern recognitionTransition

Transition

(1) Barrier of teaching rule updating

(2) Limitations of relying on logical operations

Enhanced through application-specific systems

Accumulation of platform technologies

(1) Accumulation of big data(2) Better computer performance(3) Development of statistical

techniques

General-purpose AI(Hitachi)

Machineplanning

Fig. 4—Transition for Special-purpose to General-purpose AI.AI techniques developed for specific applications are being redeployed for general application.

ICOT: Institute for New Generation Computer Technology MC: Monte Carlo


- 24 -

Learning from the Universe and EvolutionIt is anticipated, however, that searching large

information spaces, as required in level 3 and level 4

AIs, will need more than the two simplistic methods

described above.

In fact, ways of achieving this are already

becoming apparent. These methods derive from the

idea that information spaces are not uniform. In other

words, space is not an empty vessel and the outcomes

being sought are in fact very unevenly distributed.

This means there is no need to perform evenly spread

random searches of large areas.

The reason this statement can be made is due to

a deep physical principle, namely that information

is something that derives from the universe, and that

the universe itself is a form of information(13) – (15).

Approximately 14 billion years ago, immediately after

the Big Bang, the universe contained only hydrogen

and helium. Describing this state required much less

information than would be needed to describe the

present-day universe. In other words, the information

content of the universe at that time was smaller. This

was followed by numerous generations of stars forming

during which planets were also created in orbit around

these stars, including our own beautiful Earth. On

Earth, many different types of complex life arose over

a period of about four billion years, leading ultimately

to human beings. Humans in turn invented writing,

media, and other tools, subsequently giving rise to

systems, organizations, and psychology. The space that

encompasses all of these outcomes represents a huge

information space. The universe itself contains all of

the information that represents these things.

The principles that underpin the creation and

progression of this space are already known. In

physics, they are referred to as the first and second

laws of thermodynamics, namely the conservation

of energy and the rule that entropy always increases.

While these laws were initially used to explain thermal

processes such as engines and turbines, they also

govern the creation and progression of all things.

A simple way of putting it is to say that, while

resources are constant, neither increasing nor

decreasing, information is continually increasing

through new combinations of resources and therefore

the universe has come to contain a greater diversity of

information. This combines the law of conservation

of resources with that of increasing diversity of

information.

The important factor in this is that conservation

of resources and diversity of information are not

proposed that the objectives of learning machines can

only be achieved by trial and error. It is through this

trial and error that computers derive their intelligence.

The following explains in more detail.

All AIs, regardless of their level, build models for

improving outcomes from large amounts of actual

data. What this actually involves is finding, by trial and

error, the conditions in this large information space

under which the outcome is maximized.

While this may not seem like much, the availability

of large amounts of data and improvements in the speed

of computers mean that, in practice, the difficulties of

AI have become concentrated around the problem of

how to search an information space.

However, the larger the search space, the more

the search process comes to resemble prospecting for

gold. The higher the AI level (as defined above), the

greater the dimensionality and the more difficult the

search. Accordingly, the key to AI lies in how to search

such a large space of unknowns both efficiently and

systematically.

Traditionally there have been two ways to go about

such a gold prospecting search. The first is to perform

a random search. This is equivalent to saying that,

when dealing with the unknown, you may as well

search blindly.

The other method is to search in the vicinity of

places where previous searches have found good

results (have obtained a high value for the outcome).

This is based on the assumption of continuity in the

information space, such that there must be a reason

why good results are obtained in a particular location,

and therefore it can be anticipated that a “seam of

gold” must be present somewhere near places that

have delivered good results in the past.

While algorithms based on this principle go under

a wide variety of names that are prone to confuse

non-specialists, they are fundamentally quite simple.

They consist of the Monte Carlo method(g) (which

is based on random numbers) and its derivatives,

with techniques used in AI including the Markov-

Chain Monte Carlo (MCMC) method, Hamiltonian

Monte Carlo method, particle filtering, Boltzmann

machines, simulated annealing, genetic algorithms,

and quantum annealing. Deep learning neural

networks, a topical subject in recent times, use these

techniques internally.

(g) Monte Carlo method

A calculation technique for obtaining an approximate solution by

repeating a simulation many times using random numbers (random

digits like those obtained by throwing dice).


- 25 -

RELATIONSHIP BETWEEN PEOPLE AND AI

Does Technology Make People Happy?In recent times, the subject of the relationship between

AI and people has arisen frequently in books and other

media.

There is no doubt that technology has made

people’s lives easier up to now. However, the question

of whether technology makes people happy is not such

an easy one to answer.

Whether AI will make people happy is becoming

an important question.

Speaking personally, even since the author was

a university student, he has had a strong interest in

the questions of what constitutes happiness and what

can be done to bring it about. Back then, he was

particularly fond of the book, Happiness: Essays on the Meaning of Life by the Swiss philosopher Carl

Hilty(18). After starting at Hitachi, however, his focus

shifted to things like technology and money, and it did

not occur to him to consider happiness in his work.

As explained earlier, the author had to start all

over again 20 years after joining Hitachi when he

and his colleagues had their careers reset. Having

had their escape routes cut off, their thoughts about

what future direction they should now take related

to the importance of data. That is, they discussed the

idea that data, including data about people, would

be more important in the future. The author believes

that, somewhere in the background of all this, he was

influenced by his interest in people’s happiness during

his time as a student.

As a result, a decision was made to develop devices

for measuring large amounts of data about people,

leading to the development in early 2006 of prototype

wearable sensors in the form of wristbands and

nametags, the latter being worn on the chest. Confident

of the small size and low power consumption of the

devices, they could be used to collect a steady stream

of data on people powered only by a small battery.

Accelerometers were of particular interest(19), (20).

A wearable sensor with a built-in accelerometer can

collect data on bodily movements 24 hours a day.

The idea was that it would be possible to determine a

variety of information about people’s behavior from

their movements (see Fig. 5).

In the author’s role as the leader of this project, he

was its first experimental subject. Since March 16,

2006, the wristband wearable sensor has remained

on his left wrist throughout the 10 years that have

passed. This means that all of the movements of his

independent of each other, rather, they make up two

halves of the same law. The principle by which the

universe is created and expands is that of a continual

increase in entropy (diversity of information) under

the constraint of a conservation of energy (resources).

The diversity of information serves as the universe’s

objective function and the limitation on resources

serves as the constraint.

Even with finite resources, information can increase

indefinitely through combination. Here, it is helpful to

recall the formulas taught at school about permutations

and combinations. Use of factorials can quickly result

in extremely large numbers. This means that the number

of potential resource combinations is nearly infinite.

This indicates the desire expressed above relating

to searching large spaces. If information becomes

increasingly diverse without constraints, it is necessary

to search this large space uniformly. In this case,

performing needle-in-a-haystack searches becomes

increasingly difficult as the number of dimensions of

the space increases. If, on the other hand, information

becomes increasingly diverse under the constraint

of finite resources, it can be shown that the desired

outcomes will be very unevenly distributed(16). This

means there is the potential to significantly narrow

down the space to be searched.

The H general-purpose level 2 AI already makes

use of this principle in some of its operations. It is

anticipated that achieving levels 3 and 4 will require

more systematic use of the principle.

As to whether it is really possible to build level 3 or

level 4 AIs, there is no doubt about this because their

existence has already been proven by the evolution

of life.

Evolution was frequently assumed in the past to be

a random search in the form of natural selection, with

survival as the objective. However, the mathematician

Gregory Chaitin has asserted the need for a new

principle, stating that the complexity of modern

lifeforms could not have been achieved in four billion

years by a random search(17). Rather, it seems likely

that evolution is already putting the above principle

to good use.

Discussion of AI often calls on comparisons

with the brain, however, the secret key to effectively

navigating large information spaces can be found in

the mechanism of evolution, which gave rise to even

the brain among other things. This could provide

a mathematical principle for explaining how the

universe has come to be the way it is. We still need to

learn from nature.


- 26 -

interesting discovery. Individual data values only

represent wrist movements, and this data on its own

is of no value. However, it was also found that more

meaningful conclusions could be drawn when the data

was consolidated and patterns were observed.

Something occurred to the author as he looked at

and analyzed the data on a daily basis: could it be that

the data contains patterns indicative of the person’s

happiness?

Subsequently, more than a million days’ worth

of data was collected on people with millisecond

resolution. The data covered a wide range of work

in different industries. An analysis of the data, which

left wrist over the last 10 years have been recorded on

a computer(16).

The visualization technique adopted to present this

data is a graph called a life tapestry. Fig. 6 shows a

trace of the data over the last seven years. Sleeping and

waking, commuting, lunch breaks, and everything else

from going on overseas business trips to working in

the office is reflected in the data on wrist movements

and is visible at a glance.

The data was collected and presented for a large

number of users and an analysis was performed of

the correlations with the results of questionnaires

about personal activities. The results contained an

Wristband sensor

2 Hz

Acceleration along three axes

Acceleration along three axes

0.8 Hz1 second

1 second

Nametag sensor

Bodily movement Walking

E-mail

Fig. 5—Wearable Wristband and Nametag Sensors and Acceleration Waveforms (x, y, and z Axes).Walking gives a waveform with a frequency of approximately 2 Hz and writing an e-mail gives an intermittent waveform with an average frequency of about 0.8 Hz.

2009

January

February

March

April

May

June

July

August

September

October

November

December

0 0 0 0 0 0 024 24 24 24 24 24 24

2010 2011 2012

Time (hour)

2013 2014 2015

Fig. 6—Bodily Movement of Author over Past Seven Years (at 50-ms Intervals).The data is called a life tapestry. Red indicates active movement, blue indicates being stationary, and intermediate colors indicate intermediate states. Things like shifts in time zones due to overseas travel and variations in sleeping time can be seen at a glance. The reason for the increase in the variegated pattern of red during 2015 was due to a rapid increase in the number of lectures and other presentations on AI given by the author.


- 27 -

The bodily movement patterns referred to here

quantify the diversity of movement in a group. First,

bodily movements were classified based on whether or

not they were static. This totally unconscious pattern

was then used to represent people’s movements as a

pattern of 1s and 0s, as in a bar code. There were cases

where people appeared to have moved but stopped again

within one minute, and cases when they continued to

move for 20 minutes or more. The results indicated

that, in practice, once people start moving, they often

continue moving for about 10 minutes on average.

When this pattern of movement continuity was

looked at for active organizations with a high level of

happiness, what was found was a mix that included

movement of both short and long duration. This was

interpreted as a projection onto the time axis of a wide

range of actions taking place at the organization.

In organizations with a low level of happiness

and activity, on the other hand, this diversity of

movement was low. In an extreme case, this might

involve frequent instances in which people continued

to move en masse for 10 minutes or so once they

started moving, and then stopped again. There were

no drivers to sustain the movement.

The term organizational activation was used to

indicate this numeric representation of diversity in the

duration of sustained movement. It represents bodily

movement, which has a strong correlation with the

happiness of the organization.

By using this technique, it is possible to measure

organizational happiness as if it were weight or height.

Rules of Happiness in OrganizationsOnce it became possible to measure happiness,

something believed in the past to be impossible to

measure, a series of previously unnoticed regularities

were identified in organizations and work. These were

collated in the form of three rules about happiness and

are described below using the example of a call center.

The first rule is that organizations with a high level

of happiness also have a high level of productivity.

Many people will associate terms like happiness or

wellbeing with the suspicion that the staff are taking

things easy. The data clearly refutes this.

It was found at an outbound call center that,

compared to below-average days, order rates were

34% higher on days with a high level of organizational

activation, meaning a high level of happiness in the

organization (see Fig. 8). Likewise, sales at a retail

store were 15% higher on days with a high level

of organizational activation. Similar results were

included the use of an AI, succeeded in identifying

patterns of people’s happiness(21), (22). This result was

announced to the press in February 2015, and the

author wrote an article that appeared in the Japanese

edition of the Harvard Business Review.

Measuring HappinessHow is it possible to measure happiness? Hitachi

conducted a 20-question survey of 468 people across

10 organizations. How many days were you happy

this week? How many days were you feeling good,

lonely, or sad? The subjects were asked to respond to

these questions on happiness with a ranking between

zero and three. When the results were collated for

each organization, it provided a quantitative measure

of whether the organization was happy on average.

Organizations with a high level of wellbeing and activity

scored highly and those with low levels scored poorly.

The bodily movement patterns of the subjects were

also measured by wearable nametag sensors worn on

the chest. The results showed a very strong correlation

between particular figures on bodily movement

patterns and the results of the above happiness survey

(see Fig. 7). The high correlation coefficient of 0.94

indicates that the happiness of an organization can

be measured using wearable sensors alone, without

performing a survey or similar study.

Diversity of activity at organization(level of organizational activation)

R = 0.94Very high accuracy

Hap

pine

ss s

urve

y

00

5

10

15

20

5 10 15 20

Questions about happiness

20 questions about the previous weekWellbeing, happiness, loneliness, sadness, etc.

Fig. 7—Identification of Bodily Movement Patterns that Correlate with Group Happiness.The diversity of activity at an organization is strongly correlated with the results of surveys (questionnaires) of happiness at the organization (with a correlation coefficient of 0.94). The total sample size was 5 billion data points, with 5,000 person-days of data collected from 468 people across 10 organizations (statistically significant at less than 1%).


- 28 -

This is because the exact same situation occurred at the

call center and it was possible to capture quantitative

figures on this in big data.

At the outbound call center, staff members call

potential customers to sell them products. Performance

is measured by the product order rate. A look at the data

on order rates over the previous half-year shows that

the staff includes both people with high performance

(like the fourth batter in a baseball team) and others

who do not perform as well. The make up of the staff

varies from day to day, with a large number of high-

performing staff members on some days and a low

number on others. Naturally, it can be predicted that

the overall order rate for the center will be high on days

with a large number of high-performing staff members.

In practice, however, no such correlation was found.

Bringing together a large number of high

performers such as number-four batters does not create

a strong team. A degree of group effect is present even

in an individual activity like making calls and taking

orders as instructed by a manual. The effect is likely

to be greater still in work that requires team players.

Together with this, performance was 34% higher on

days when there were a large number of people with

a high level of diversity in their bodily movements

and people who increase the diversity of bodily

movements of those around them. It seems likely that

these are people who improve the workplace mood.

At a workplace where great effort is made to improve

the order rate by even 1%, a 34% improvement is very

large indeed.

What is interesting when looking at the performance

of those people who contribute to the mood of the

workplace is that there is no correlation with order rate.

obtained from development projects. Measurements

were taken in four projects and it was found that the

subsequent sales contribution from the deliverables

developed was higher in those projects with a high

level of organizational activation in the second month

after the project started. Projects with a high level

of happiness delivered better financial results. Put

another way, bodily movement predicted the success or

otherwise of a project long before the financial results

became apparent. Using this knowledge, action can

be taken on a project earlier than is currently the case.

The second rule is that both happiness and financial

performance are group phenomena. There is a tendency

to think of happiness as something that exists in the

minds of individuals. The data clearly refutes this.

The following words were spoken by Kimiyasu

Kudo of the Fukuoka Softbank Hawks professional

baseball team after a victory.

When we are losing and I ask the team what’s

happened to their voices, people like Kawashima

and Fukuda (motivators) yell for us to go for it.

When everyone shouts it together we really do go

for it. Creating the right sort of atmosphere on the

bench is also important.

(Sports Nippon, September 18, 2015)

(Author’s translation)

This quote is saying that the atmosphere on the

bench is very important for winning, to which the

bench players Kawashima and Fukuda made a major

contribution. While this may look at first glance like

a belated expression of gratitude to the bench players

after a win, I believe it was just as Kimiyasu Kudo said.

Call center Retail store Development projects

34%

15%

0.8 12

12

PJ1 PJ2

PJ3

PJ4

13 14 17

8

4

0

2,200

2,000

1,800

1,600

1,400

1,200

1,000

0.5

0Low

Ord

er r

ate

Sale

s pe

r cu

stom

er (

yen)

Five

-yea

r sa

les

(arb

itrar

y sc

ale)

High Low High

Organizational happiness = Diversity of activity(average for each day)

Organizational happiness = Diversity of activity(average for each day)

Organizational happiness = Diversity of activity

(second month after start of project)

Fig. 8—Relationship (Level of Organizational Activation) between Happiness and Productivity.Examples are shown for a call center, retail store, and projects. The call center and store both achieved good results on days with a high level of movement diversity (level of organizational activation). The level of organizational activation in four development projects in the second month after starting was correlated with the contribution to sales of the development outcomes (statistically significant at less than 5%).


- 29 -

For example, happiness dashboard software was

developed for the call center that outputs instructions

to the supervisor as to who they should focus on talking

to on that day(25). The software was used for more than

one year and improved the order rate by 27%.

Happiness is InfectiousThis work has been publicized since the spring of 2015

and has generated a considerable response, including

the publication of news releases describing work at a

bank and an airline(26), (27).

Of course concepts like happiness and wellbeing

are not things that were invented or proposed by the

author. In fact, there has also been growing interest

in happiness in academia over the last 15 or so years.

It has been reported that people with a high level of

happiness are more likely to enjoy good health, a long

life, and successful marriage, with better income and

career performance, 37% higher sales productivity,

and 300% higher creativity, and that companies with

a large number of happy people have 18% higher

earnings per share than companies with few such

people. Papers like this are appearing on a monthly

basis(28).

However, happiness in the past has been measured

by questionnaires, and given the inability to take

realtime measurements, there has been little clarity

as to what actions serve to improve happiness. Along

with the divide between academia and practical

business, this has led to little use being made of

these findings in business. This situation has been

transformed by the use of wearable sensors for

measurement and feedback using general-purpose

AI. The prerequisites for utilizing this knowledge in

business are at last in place.

In practice, small things can influence happiness.

Hitachi conducted experiments jointly with Professor

Sonja Lyubomirsky and Dr. Joseph Chancellor of

the University of California, Riverside(29). A group

was divided randomly and the experimental subjects

were asked to write down three good things that had

happened to them during the past week. Similarly,

the control subjects were asked to write down three

things that had happened to them during the past

week. The only difference was the inclusion of the

word “good.” This was repeated for five weeks and

the results were reviewed for differences two months

later. Clear differences were evident in the results.

The experimental subjects had higher happiness, a

greater sense of belonging to their organization, and

a higher level of bodily activity from the morning.

These people improve the results of those around them

by creating a better atmosphere around them. Don’t all

workplaces have people like Kawashima and Fukuda?

But it is likely that people like this are not necessarily

recognized by existing human resource practices.

The data also clearly shows that communication

has a large influence on generating such a workplace

mood. It was found that both happiness and order

rates at the call center were high on days when there

was vibrant conversation during work breaks. It was

also found that both happiness and order rates were

strongly influenced by who the supervisor spoke to

on that day(23).

The third rule is that happiness is represented by

a single measure regardless of the work or people

involved. Many people will imagine that the definition

of happiness will itself vary depending on things like

the company, type of work, and region. However, the

high 0.94 correlation coefficient does not allow for

this interpretation. It means that a single measure can

be used to represent happiness.

It follows that AIs can be used to improve

happiness, and thereby to also improve productivity.

Specifically, Hitachi has developed a system that

inputs staff activity data captured by wearable sensors

into the H general-purpose AI and provides feedback

on the communications and uses of time needed to

improve the happiness of those individuals and those

around them(24) (see Fig. 9). This has enabled the AI

to understand people’s happiness.

Fig. 9—Photo of the Happiness Improvement Support Service Utilizing Wearable Sensors and AI.By having staff wear wearable nametag sensors, the service is able to provide feedback on the communications and uses of time that will enhance the level of organizational activation (which is correlated with productivity and happiness). Feedback provided using AI is customized automatically to the circumstances in the user organization.


- 30 -

this problem (see Fig. 10). One approach is professional

services, which include a wide range of services such

as consulting, law, and accounting. These services

are provided in the form of work, making them labor-

intensive. While some individuals may be capable

of high performance, the larger the organization, the

less variation there tends to be between companies on

average. To achieve profits, such services often use

star players to front their business but have the actual

work done by young staff on lower wages.

The other approach involves process-management-

based services. These are services in which processes

can be standardized, such as distribution warehouses,

call centers, and equipment maintenance. Costs are cut

and profits achieved by using manuals and training.

Services such as IT lie midway between these two

approaches.

AI overcomes this tradeoff between individual

optimization and cost to enable scalable services.

Advertising services that are linked to Google*2 or

other searches already earn margins that are an order of

magnitude greater than traditional advertising services.

These search-linked advertising services use a particular

type of special-purpose AI to deliver advertising services

that are customized to specific customers. Moreover,

they improve customer value without incurring costs

thanks to fully automatic customization. This enables

businesses that transcend the above tradeoff.

It is anticipated that this model can be applied to

a wide range of services by using general-purpose

AI. Services that use AI are increasingly automating

customization for specific customers.

Another important point is that the use of AI

makes services more outcome-oriented. In the past,

suppliers were only able to offer their own products

and associated services. A supplier of production

machinery, for example, would also have supplied

repair and other maintenance services. However,

customer issues have a hierarchical structure. This

hierarchy extends from top-level issues that directly

influence the outcome of customer profits to more

indirect low-level issues(30). Wireless communication

operators, for example, face the issues of operating

their network successfully and attracting customers that

influence profit directly, and also low-level issues such

as the appropriate construction of the physical layer.

While higher profits can be achieved if solutions can

be provided to top-level issues that influence outcomes

directly, the delivery of higher level solutions is difficult

We were surprised that this action, which took only

about five minutes each week to perform, could

produce such a difference. It demonstrated how the

subjective phenomenon of happiness could be altered

significantly by small changes. There is still room for

improvement in people.

HOW AI WILL CHANGE SERVICES

AI has the disruptive power to change services.

In advanced economies such as Japan, the heart

of the economy is shifting from manufacturing to

services. Whether its previously strong manufacturers

can become service-oriented is a major issue for Japan.

There are common challenges faced by anyone

seeking to improve profits from services. The value

of services is enhanced by solving the problems

of customers. However, individual customers have

different issues. Value is enhanced when solutions

can be offered that are tailored to these individual

issues. On the other hand, there is a cost associated

with trying to supply individually tailored solutions.

There is a tradeoff between these costs and the value

delivered to customers, a situation in which profits are

difficult to achieve.

Accordingly, many service industries have low

profitability on average. Industry-average profit

margins in areas such as distribution, retail, advertising,

hospitality, and accommodation are below 5%.

In the past, services have adopted two approaches to *2: Google is a trademark of Google Inc.

Professional services• Consulting• Legal advice• Accounting and auditing• Architecture• Design• Advertising agencies

IT services• Requirements definition

• Schematic design• Detailed design• Testing and QA

Process-management-based services• Equipment maintenance services• Logistics• Call centers• Facilities management• Business outsourcing

Specialist services

Process-management-based

Cost of service delivery

Cus

tom

er is

sues

to

be s

olve

dH

igh-

leve

l cus

tom

er is

sues

High cost Low cost

Low

-leve

l cus

tom

er is

sues

AI-based services

Fig. 10—Tradeoff Between Customer Value and Cost in Services.The use of AI enables high-level issues that derive from customer outcomes to be supplied at low cost.


- 31 -

to be proven applications, if different data is used, the

results will be completely different. Customers also

need to be made to understand this possibility.

There is a need to reform staffs and organizations

to provide opportunities for training and practice so

that these things can be utilized in practical marketing.

Reform of Integration OrganizationsThe second is the reform of integration organizations.

There is a need for flexibility that goes beyond the past

practice of system integration (SI) driven by customer

specifications.

General-purpose AI can transform the mechanistic

IT and other equipment of the past into IT and

other equipment that can learn and grow. Infinite

possibilities are made available by combining this

general-purpose AI with conventional IT and other

equipment, and it has the power to generate significant

value for customers. These IT and other equipment

systems do not distinguish either between in-house

and outsourced products, or between existing and new

installations. Value can be generated through diverse

combinations of AI with these in-house, outsourced,

existing, and new installations.

To begin with, the combination of AI with existing

IT and other equipment opens up opportunities for

new services. In the past, maintenance services for IT

and other equipment have been an important business.

An ongoing and reliable source of income can be

created by augmenting an equipment business with

maintenance services. General-purpose AI takes this a

step further. It is possible to achieve ongoing increases

in asset value by integrating a general-purpose AI with

existing equipment and IT. This is because it enables a

steady progression of new management and business

challenges to be dealt with in an ongoing manner by

changing the configuration.

In the distribution warehouse example mentioned

above, scheduling functions able to adapt flexibly

to fluctuations in supply and demand or variations

between staff were added to an existing WMS by

connecting it to a general-purpose AI. New functions

that in the past would have needed to be added

to the WMS were instead achieved by adding a

general-purpose AI. Furthermore, daily changes in

circumstances are identified automatically from data

and control logic that are updated automatically. Once

the connection is in place, the optimal allocation of

a wide range of people, goods, and money related to

WMS data can be achieved simply by changing the

AI configuration. Ongoing improvements to system

when limited only to the company’s own products.

This is a situation where the benefits of a general-

purpose AI can be utilized. If a general-purpose AI can

be connected to existing systems to obtain input data,

outputs with value to the customer can be obtained

from customer outcomes. To achieve this, a company

must not be bound by its own products during system

implementation. There is greater potential for winning

large orders through the orchestration of a wide range

of products, including those from other suppliers.

CORPORATE ORGANIZATION IN THE ERA OF AI

AI enables services that generate ongoing innovation,

with businesses that use AI being outcome-oriented.

Accordingly, they require viewpoints and actions

different to those of traditional product or service

organizations. Organizations will need to undertake

the following three reforms.

Reform of Frontline OrganizationsThe first is the reform of frontline organizations. There

is a need for frontline staff members and organizations

that can engage with and market to customers in ways

that derive from outcomes.

The concepts associated with this are significantly

different from those of traditional product-based

negotiation and marketing capabilities. Two externally

evident differences are the need to ask customers

about the outcomes they want and what outcomes to

use to evaluate the success of the solutions to these

challenges. There is also a need to discuss the value

of suggested outcomes and how this value relates to

their associated investment and other costs. This will

require consultative abilities to avoid outcomes that

are not likely to provide a cost-benefit and choosing

outcomes that are likely to be beneficial.

Furthermore, existing sales staff members and

system engineers will initially have no appreciation

for the flexibility of general-purpose AI. Accordingly,

even after they have succeeded in identifying new

requirements from customers, they will tend to be

hesitant about adopting unproven methods. This is

because they are used to the rigid view of products

and marketing practices of the past.

The revolutionary feature of general-purpose AI is

that it can be used for many different functions simply

by changing the input configuration. The value of this

will be wasted if the technology is limited only to

proven applications. Conversely, even in what appears


- 32 -

to incorporate the subsequent new requirements into

the platform. This will require the establishment of

practices that are capable of high rotation in a more

agile and DevOps(h) manner than past low-level

platforms(31).

The aim of the above three reforms is simple. It

is to transform organizations into ones that generate

innovation. According to Peter Drucker, the purpose

of companies is to create customers. Innovation is the

act of generating new demand for this purpose.

Example of Organizational Reform: Reform of HitachiThe new business structure adopted by Hitachi in

April 2016 provides an example of the form taken by

organizations for services that use AI(32) (see Fig. 11).

In place of the previous product-oriented company

structure, it is made up of customer-driven frontline

organizations together with horizontal service and

platform organizations and product organizations.

For the frontline organizations, the aim is to create a

business structure that generates innovation alongside

customers by setting up 12 business units (BUs) to

improve frontline functions such as sales, engineering,

and consulting in four markets: (1) Power/Energy, (2)

Manufacturing/Water, (3) Urban Development, and (4)

Finance/Public/Healthcare. Hitachi also set up service

and platform BUs that consolidate and integrate the

technologies essential to providing advanced services,

including AI, analytics, and control technology, to

create a business structure that supplies open common

platforms to the frontline BUs and other partners.

The product-focused businesses supply components,

materials, and other products to the frontline BUs and

other customers.

In this way, moves to reform organizational

structures can be expected to become more widespread

in the future along with the emergence of the innovative

technology of AI.

CONCLUSIONS

This article has presented an overview of AI, which

has a potential that goes beyond just advances in

computing. As a new methodology for humankind,

AI dramatically enhances the ability of companies and

people to respond amid the unpredictable changes in

functions can be made in accordance with operational

and managerial requirements.

Here, the connection of a general-purpose AI is of

great significance for the business regardless of whether

it is an in-house or outsourced product. This is because

it expands opportunities for working with customers

by using IT and other equipment from other suppliers.

Consider also the significant expansion in the

scope of products that can be offered when marketing

new IT and other equipment. It is possible to offer

new systems that learn about and react to changes in

circumstances by combining a general-purpose AI

with in-house and outsourced IT and other equipment

products. Achieving this requires the establishment of

an ecosystem or other network that includes products

from other suppliers as well as in-house products.

For the new solutions using general-purpose AI

described above, both staff members and organizations

need the flexibility to combine a wide range of products

and AI. Supporting this requires an ecosystem for a

diverse range of companies capable of supplying an

extensive range of products and services to exchange

information and facilitate orchestration across products.

Reform of Platform OrganizationsThe third is the reform of platform organizations. There

is also a need to reform platform organizations in order

to enable the first and second activities described above.

The first requirement is to raise the level of technical

hierarchies handled by platforms. The platforms for

the ongoing improvement and operation of general-

purpose AIs and their configurations will handle

management outcomes at a much higher level than the

servers, storage, databases, operating systems (OSs),

and operational middleware platforms used in the

past. There is a need to change the mindset of people

who have performed past hardware, software, and

other development in accordance with specifications.

Moreover, this means that customer outcomes will

become easier to understand, including for the people

responsible for platforms. They will serve in roles where

they deal more directly with customers. It also means

being able to contact people with greater decision-

making authority. In both cases, it will put them in

situations that are more rewarding. One obstacle to this

is anxiety and fear of change. This will likely need to

be dealt with in an organizational manner.

Dealing with higher level matters means greater

exposure to change. It requires an organization in

which 100 new AI applications each month commence

operation using new configurations and that is able

(h) DevOps

A term coined from the combination of development and operations.

It is a method for developing systems that adapt more quickly and

flexibly to business challenges through cooperation and collaboration

between the people responsible for development and operations.


- 33 -

The poem expresses, with a grand sense of scale,

the feeling of contemplating the unknown world

prompted by the trials of a disability.

In acquiring AI as a new way of confronting the

unknown, can people investigate new things every

day? People want to progress by routinely making new

investigations into the unknowns faced in corporate

strategy, business operations, national policy making,

and the lives of individuals. This approach by people is

likely to be always creative in facing the possibilities

of the unknown while also being incomplete. The

author would like to be like this.

Completion consists of being forever incomplete

(Kenji Miyazawa 1926(34)) (Author’s translation)

REFERENCES(1) D. Silver et al., “Mastering the Game of Go with Deep Neural

Networks and Tree Search,” Nature, Vol. 529, pp. 484–489

(Jan. 2016).

(2) P. F. Drucker, “Management Challenges for the 21st Century,”

HarperBusiness, New York (1999).

the environment that are occurring globally. The power

of this new way of doing things will grow rapidly over

time compared to previous practices that relied on

the capabilities of individuals. This trend will extend

across all industries and types of work. The author

refers to this as the general adoption of AI.

Nevertheless, AI is a methodology for humankind.

Its purpose is to increase the ability of people to solve

problems and it is the task of people to decide which

problems to solve, a process that calls heavily on

experience and intuition. This will enable people to

take on challenges they could never have imagined in

an unpredictable world. The creation of the unknown

future is something we can start today.

After being paralyzed on her left side by a stroke at

the age of 77, the sociologist Kazuko Tsurumi wrote

a poem about encountering a world she had been

unaware of when healthy.

Small universe, named I, resonates in the large

universe

The sound it makes is new every day(33)

(Author’s translation)

Group Corporate

Power/Energy Manufacturing/Water

UrbanDevelopment

Finance/Public/Healthcare

Services & Platforms BU

Industrial Products BU Group companies

Products, components, materials, etc.

Product-focused Businesses

Service-focused Businesses

President & CEO

12 Front BUs

Ener

gy S

olu

tions

BU

Urb

an S

olut

ions

BU

Pow

er B

U

Wat

er B

U

Bui

ldin

g Sy

stem

s B

U

Def

ense

Sys

tem

s B

U

Rai

lway

Sys

tem

s B

U

Reg

iona

l Chi

ef E

xecu

tive

for A

sia-

Paci

fic

Reg

iona

l Chi

ef E

xecu

tive

for

the

Am

eric

as

Reg

iona

l Chi

ef E

xecu

tive

for

Chi

na

Reg

iona

l Chi

ef E

xecu

tive

for

EM

EA

-CIS

Fina

ncia

l Ins

titut

ions

BU

Gov

ernm

ents

& P

ublic

Cor

pora

tions

BU

Hea

lthca

re B

U

Indu

stry

& D

istr

ibut

ion

BU

Nuc

lear

Ene

rgy

BU

Research & Development Group

CEO: chief executive officer BU: business unit EMEA: Europe, the Middle East, and Africa

CIS: Commonwealth of Independent States

Fig. 11—Example Organization that Uses AI (Reorganization of Hitachi).AI plays an important role in Hitachi’s new structure made up of customer-driven frontline organizations together with horizontal service and platform organizations and product organizations. The reorganization took place in April 2016.


- 34 -

(23) J. Watanabe et al., “Resting Time Activeness Determines

Team Performance in Call Centers,” ASE/IEEE Social

Informatics (Dec. 2012) pp. 26–31.

(24) S. Tsuji et al., “Use of Human Big Data to Help Improve

Productivity in Service Businesse,” Hitachi Review 65,

pp. 847–852 (Mar. 2016).

(25) J. Watanabe et al., “Workscape Explorer: Using Group

Dynamics to Improve Performance,” CHI ‘14, Ext. Abstracts

(2014) pp. 2209–2214.

(26) Hitachi News Release, “Assisting Efforts to Improve Work

Productivity at Bank of Tokyo–Mitsubishi UFJ through Big

Data Analytics Utilizing Artificial Intelligence Technology”

(Sep. 2015).

(27) Japan Airlines Co., Ltd. and Hitachi, Ltd. joint press release,

“JAL and Hitachi Launch Joint Demonstration Project

Aiming to Improve Employee Satisfaction Utilizing the IoT

and Artificial Intelligence” (Oct. 2015).

(28) Special Issue, The Value of Happiness: How Employee Well-

Being Drives Profits, Harvard Business Review, Jan.-Feb.

(2012).

(29) K. Yano et al., “Sensing Happiness: Can Technology Make

You Happy?” IEEE Spectrum (Dec. 2012) pp. 26-31.

(30) M. Imaeda, “Two Different Approaches to Service

Management: Advice for Manufacturers,” Hitotsubashi

Business Review, 54 (2) 2006AUT (Sep. 2006) in Japanese.

(31) M. E. Porter and J. E. Heppelmann, “How Smart, Connected

Products Are Transforming Companies,” Harvard Business

Review (Oct. 2015).

(32) Hitachi News Release, “Hitachi to Make a Transition to

a Market-Specific Business Structure with Strengthened

Front-Line Functions—Providing innovations through a

combination of services and products” (Feb. 2016).

(33) K. Tsurumi, “A collection of poems: Yamauba (Dame of the

Mountain),” Fujiwara-shoten (Jul. 2001).

(34) K. Miyazawa, “An Outline Survey of Peasant Art,” the

Complete Works of Kenji Miyazawa 10, Chikumashobo Ltd.

(1995).

(3) P. F. Drucker, “Managing for Results,” Harper & Row, New

York (1964).

(4) R. Kurzweil, “The Singularity is Near,” Loretta Barrett Books

Inc. (2005).

(5) N. Bostrom, “Superintelligence: Paths, Dangers, Strategies,”

Oxford University Press (2014).

(6) M. Shanahan, “The Technological Singularity,” The MIT

Press (2015).

(7) J. Barrat, “Our Final Invention: Artificial Intelligence and the

End of the Human Era,” Griffin (2015).

(8) G. E. Hinton and R. R. Salakhutdinov, “Reducing the

Dimensionality of Data with Neural Networks,” Science 313,

pp. 504–507 (Jun. 2006).

(9) S. Baker, “Final Jeopardy: The Story of Watson, the Computer

That Will Transform Our World,” Mariner Books (Mar. 2012).

(10) Q. V. Le et al., “Building High-level Features Using Large

Scale Unsupervised Learning,” Proc. of the 29th International

Conference on Machine Learning (2012).

(11) E. Brynjolfsson, A. McAfee, “The Second Machine Age:

Work, Progress, and Prosperity in a Time of Brilliant

Technologies,” W. W. Norton & Company, Inc. (2014).

(12) A. M. Turing, “Computing Machinery and Intelligence,”

Mind 59, pp. 433–460 (1950).

(13) E. T. Jaynes, “Probability Theory: The Logic of Science,”

Cambridge University Press (2003).

(14) H. C. von Baeyer, “Information: The New Language of

Science,” Weidenfield & Nicolson Ltd. (2003).

(15) S. Lloyd, “Programming the Universe: A Quantum Computer

Scientist Takes on the Cosmos,” Vintage (2007).

(16) K. Yano, “Invisible Hand of Data: The Rule for People,

Organizations, and Society Uncovered by Wearable Sensors,”

Soshisha Publishing Co., Ltd. (Jul. 2014) in Japanese.

(17) G. Chaitin, “Proving Darwin: Making Biology Mathematical,”

Pantheon (May 2012).

(18) C. Hilty, “Happiness: Essays on the Meaning of Life,”

translated by Francis Greenwood Peabody, Nabu Press

(1891).

(19) T. Tanaka et al., “Life Microscope: Continuous Daily Activity

Recording System with a Tiny Wireless Sensor,” Proc. 5th Int.

Conf. Networked Sensing Systems (INSS 2008), pp. 162–165

(2008).

(20) Y. Wakisaka et al., “Beam-scan Sensor Node: Reliable

Sensing of Human Interactions in Organization,” Proc. 6th

Int. Conf. Networked Sensing Systems (INSS 2009), pp. 58–

61 (2009).

(21) K. Yano, “Happiness Can be Measured by Wearable Sensors:

Enhancing Productivity in the Office through the ‘Invisible

Hand of Data’,” Harvard Business Review (Japanese Edition),

pp. 50–61 (Mar. 2015) in Japanese.

(22) K. Yano et al., “Profiting from IoT: The Key is Very-large-

scale Happiness Integration,” 2015 Symposium on VLSI

Technology, pp. C24-C27 (Jun. 2015).

Kazuo Yano, Dr. Eng.Research & Development Group, Hitachi, Ltd. He is currently engaged in research and development work including artificial intelligence in his role as Corporate Chief Scientist. Dr. Yano is a Fellow of the IEEE and member of the Institute of Electronics, Information and Communication Engineers (IEICE), The Japan Society of Applied Physics (JSAP), The Physical Society of Japan (JPS), and The Japanese Society for Artificial Intelligence (JSAI). Author of “Invisible Hand of Data,” Soshisha.

ABOUT THE AUTHORS


- 35 -

Featured Articles

AI Technology

Achieving General-Purpose AI that Can Learn and Make Decisions for Itself

Norihiko Moriwaki, Ph.D.

Tomoaki Akitomi

Fumiya Kudo

Ryuji Mine

Toshio Moriya, Ph.D.


OVERVIEW: There are increasing trends toward utilizing the big data collected at companies and the data acquired through the IoT to generate new management value. Of these trends, AI technology is attracting attention as a means of intelligently utilizing large amounts of data by making full use of computer processing capacity. Hitachi is developing Hitachi AI Technology/H, which optimizes and automates decision-making in various ways through the utilization of large amounts of data, thereby contributing to improvements in outcomes for corporations. This article first describes the trends in and types of AI technologies, and then explains the concepts and basic principles of Hitachi AI Technology/H Hitachi has developed. Furthermore, it touches on the technology’s potential as a general-purpose AI solution that can shed light on complicated problems through possible applications in a wide range of different industries and fields.

INTRODUCTION

MOVEMENTS are picking up steam in every industry towards the utilization of the large amounts of big data being collected by companies and the new data being gathered through the Internet of Things (IoT) to enable management reforms such as internal operational reforms and the creation of new customer-oriented services. As data volumes increase, however, the requirements of traditional methods based on having humans test hypotheses exceed the limitations of human cognition, making it difficult to utilize data effectively. This makes it necessary to use artificial intelligence (AI) that can automate the extraction of value from large quantities of data by applying intellectual algorithms that can replace humans(1). Expectations are also high for AI that can learn by itself in order to acquire knowledge.

This article describes the concepts and basic principles of Hitachi AI Technology/H(2) promoted by Hitachi (hereafter referred to as H), and the possibilities for application in a diverse range of industries and fields.

AI: A NEW WAY TO UTILIZE COMPUTERS

Toward Computers that Can Learn from DataThe general roles played by traditional computers

and business applications were to execute predefined functions and automate business processes. In other words, the flow of development traditionally entails designers (humans) envisioning and designing functions in advance, after which programmers code the logic. As the performance of computers continues to increase and prices continue to fall, this is expected to enable more intellectual processes in the future as opposed to the traditional fixed processes. Specifically, the availability of a wide variety of different types of data, including the IoT and other types of sensing, technological advances such as the expansion of feedback targets including wearable devices and humanoid robots, and other factors are working together to help facilitate flexible responses from computers to changes in data, thereby enabling learning from experience (see Fig. 1). When it comes to intellectual computer processing, although intense research efforts have been focused on the field of AI since the 1980s, the type of AI for which humans build in cause and effect rules in advance has not been able to understand the diverse range of user contexts, and thus has not reached the level of practical application. In recent years, however, the availability of inexpensive high-performance computers and high-capacity data storage is triggering a technological upheaval, and expectations are running high for machine-learning AI that can learn from data.

114 Achieving General-Purpose AI that Can Learn and Make Decisions for Itself

- 36 -

New AI for Optimization and Decision-makingTable 1 shows the H technology Hitachi is developing and promoting along with other representative types of AI (and intelligent systems), categorized by type. Based on the role it plays, AI can be categorized as either the searching type, which supports humans as an intelligent machine with expert systems and other functions at its core, or as the recognition type, which emulates the ability of humans to recognize images or voice input through seeing and understanding or hearing and understanding.

International Business Machines Corporation’s (IBM’s) Watson*1 and other AI systems of the

searching type take newspapers, technical papers, and other documents and text data as input and use natural language processing technology to search for information and provide responses. As for the recognition type, deep learning and other technologies have been starting to advance at a rapid rate in recent years, toward the achievement of recognition functions based on the effective extraction of patterns from massive amounts of image and voice data. Hitachi’s H goes beyond those two categories, however, by achieving automation of optimization and decision-making. More specifically, H is characterized by the ability to automatically generate models for improving specified outcomes based on a wide range of mixed numerical data types.

CONCEPTS AND BASIC PRINCIPLES OF H

Automating the Generation of HypothesesIn addition to technical challenges such as the achievement of raw computing power and storage capacity, the effective utilization of large amounts of data is also accompanied by a redefinition of the roles of both humans and computers. In other words, with the traditional approach of having humans first come up with hypotheses, gather the data that is required, and then attempt to validate the hypotheses, it is difficult to fully utilize the large amounts of frequently updated data. Also, the problems that must be solved for society or for companies are themselves growing more complicated, to the extent that even for experts in the relevant fields, the capacity of humans to recognize methods for constructing sophisticated predictive models is being exceeded. In the future, an effective approach will be to fully utilize data and computers by having humans set the problems to be solved (the outcomes to improve) so that the computers can automatically generate a large number of hypotheses and discover solutions by following the data.

Principles of HH is an analytical engine that recursively derives from large amounts of data the correct measures and how they should be implemented in order to improve the target indicators that represent value for customers [key performance indicators (KPI), etc.]. Rather than collecting models of detailed work processes, H focuses on modeling outcomes in a data-driven system.

The principles of H are shown in Fig. 2. H uses a numerical table format to input data that might influence outcomes such as management effects to

Program

Input

Instructions HypothesesAutomatic learning Facts

Input

Output

Computer

(Fixed)

AI

(Flexible)

• Sensors

• Cameras

• Robots

• Wearable devices

Output

Data

Traditional processes Future processes

Fig. 1—Computers Transitioning to AI that Learns from Data.Instead of fixed processing that must be designed in advance, this AI enables processing that is more intellectual and flexible.

Type Search Recognition Optimization and decision-making

Representative example

IBMWatson

Google*

Deep LearningHitachi

H

Data

Documents/Research

papers(Text)

Images/voice(Signal

waveforms)

Corporate information/

sensors(Heterogeneous mixed numerical

values)

Utilization situations

• Information searching

• Physician assist

• Security• Wearable user

interfaces• Creating profit

Destructive technology

Page ranks(1998, L. Page)

Deep learning(2006, G. Hinton)

Leap learning(2014, Hitachi)

IBM: International Business Machines Corporation* Google is a trademark of Google Inc.

TABLE 1. Positioning of HThe strength of Hitachi’s H technology lies in its ability to convert a wide range of mixed numerical input to corporate profit.

*1 IBM and Watson are trademarks of International Business Machines Corporation, registered in many jurisdictions worldwide.


- 37 -

be increased or decreased (sales, productivity, design

bugs, service disengagement, and so on).

H internally generates an exhaustive number of

combinations of the input data, generates huge amounts

of feature quantities for each combination, and then

uses a brute force method to calculate the relationships

between these feature quantities and the outcome in

order to discover any complicated correlations that

are latent in the data through statistical processing.

The output of H is an equation that describes the

correlation between the outcome and the feature

values of the combinations. By taking this equation

as an optimization function, and incorporating it

into work and control systems in combination with

outcome improvement prototype designs and means

of providing execution, it is possible to continuously

improve outcomes while tracking data even if the

environment or orders change.

Also, since it is easy to envision combinatorial

explosions occurring if huge numbers of data

parameters are input, a variety of different measures

are taken to inhibit this problem. One of the techniques

that was developed is spatial deformable clustering

(SDC), which is an advanced version of grid-based

clustering that enables the automatic discovery of

regions of data combinations that might strongly affect

the outcome (see Fig. 3)(3). By increasing or decreasing

phenomena with respect to the regions identified by

this technique, it is possible to effectively control the

outcome.

AUTOMATION OF SYSTEM OPTIMIZATION AND DECISION-MAKING

Online AIH is not only utilized for data analysis, but it can

also be connected to existing systems and utilized for

optimization and decision-making as well (see Fig. 4).

In other words, it regenerates models by adapting

to daily changes in the environment and orders, and

can maximize management effects by contributing

to productivity and reducing costs. Specifically, H

enables the implementation of adaptive and self-

improving systems by using equations related to the

output of outcomes as optimization functions and

narrowing down a variety of different complicated

work and control systems into mathematical

optimization problems. For example, H can be used

to automatically track changing on-site conditions in

order to improve outcomes in various environments

including manufacturing lines, warehouses, and stores.

Outcome

Features of H Rather than creating detailed models of work and systems, H focuses on modeling business results.

Basic principlesUse a brute force method to calculate correlative relationships in data, thereby discovering complicated correlations that are latent in the data (leap learning).

Exceptions• Situations that cannot be expressed as a numerical table (language, etc.)

• Situations where not enough data can be obtained

Automaticinterpreta-

tion

Several hundred thousand feature values

EquationBusiness

results = f(x)

Optimization measures

• Work data• System logs• Human actions

Fig. 2—Principles of H.Hitachi’s proprietary leap learning technology derives the factors required for improving business outcomes from the data.

4.0

3.0

2.0

1.0

8:00 10:00 12:00 14:00 16:00 9:00X 1

X 2

X 1

y Outcome

1.02.0

3.0

10:20 11:50 12:50

• Manual setting of data partitioning parameters

• Combinatorial explosion

Data combination regions are automatically discovered using automatic grid-based clustering.

1=(11:50≤ 1≤12:50) (1.0≤ 2≤3.0)z x x

Fig. 3—Spatial Deformable Clustering (SDC).The grid-based clustering technique is utilized in order to automatically discover data combination regions that might strongly affect the outcome.

H OnlineH

• Data preparation• Optimization consideration

Utilized by analysts as a tool

Optimization and decision-making services

provided.

Fig. 4—Online AI Providing Optimization and Decision-Making Services.H is connected to the system in order to automate optimization and decision-making.

116 Achieving General-Purpose AI that Can Learn and Make Decisions for Itself

- 38 -

AI that Learns from ExperienceFurthermore, H can not only be applied to past data,

it also attempts to actively search for methods that

can more optimally achieve objectives by acquiring

new data by itself. The example shown in Fig. 5 is a

prototype of an experimental system connecting H to

a swing robot that was created using the educational

LEGO*2 Mindstorms*2 product. This experiment was

designed to see whether or not the robot could learn

how to swing on a swing without being given a model

beforehand.

With the target indicator corresponding to the

outcome set as increments in the swing’s deflection

angle, and with the deflection angle, pumping state,

and pumping operation acquired from the system set

as explanatory variables, the AI gradually generates

a model by maximizing the outcome with random

pumping patterns, thereby enabling the robot to

successfully swing on the swing after an average

of approximately five minutes. Furthermore, the

robot was able to increase the outcome (the swing’s

deflection angle) by using the unexpected method of

pumping at both ends of the swing.

Although it is necessary to fully consider

possibilities such as a system running amok when

implementing enhanced learning AI of this type, it

can be used to greatly contribute in areas such as

optimizing the parameter control of complicated

control systems, or implementing robust systems that

handle changing environments.

POSSIBILITIES AS A GENERAL-PURPOSE TECHNOLOGY

The effectiveness of H has been confirmed via

application in 24 case studies covering seven fields.

The core algorithms of H are characterized by

independence from modeling or tuning based on

domain knowledge in each field. In other words,

unlike previous types of AI that had to be specialized

for each field, H is versatile enough to automate the

optimization and decision-making of a wide range of

different systems, so all that is necessary is to specify

the outcome to be maximized and to input data that

might be associated with changes in that outcome.

By taking advantage of this characteristic, H can be

applied to a diverse range of problems in society as

well as different industries, fields, and sectors. Table 2

shows the fields Hitachi is currently envisioning for

the application of H.

CONCLUSIONS

This article discussed the concepts and basic principles

of the H optimization and decision-making AI system

championed by Hitachi, as well as the possibilities of

adding intelligence to existing systems.

Whereas the type of AI that specializes in a specific

field by utilizing domain knowledge has traditionally

been in the mainstream, H’s key characteristic is

its high level of versatility. All that is necessary to

use H is to specify the outcome to be increased and

to input the data that might affect changes in that

outcome in order to automate system optimization

and decision-making. H can contribute to the solution

of a large number of complicated problems in order

to improve corporate outcomes, including the control

of parameters for manufacturing equipment, the

optimization of operation control, the proposal of

optimal products to customers, and so on.*2 LEGO and Mindstorms are trademarks of the LEGO Group.

On-site

equipmentH

Robots

Systems

Real worldIncorporate random control and

acquisition of pumping method

through trial and error

Control

standing/sitting

(pumping)

Observe

amplitude and

orientation

Active learning LEGO swing example

Optimal control of

targets

H

Formulation of search

plan

Enhanced learning

Control

Observation

Fig. 5—Active Learning.H incorporates enhanced learning capabilities in order to enable the automatic learning of optimization parameters.

Industry or field Targets of optimization and decision-making

Production Control manufacturing equipment parameters

TrafficFuel-efficient driving control, maintenance

optimization

DistributionOptimize product placement, warehouses, and

task sequences

MarketingCustomer analysis, proposal of recommended

products

Office work Automate approvals, improve happiness

TABLE 2. Fields where H is Utilized

H is highly versatile because its level of dependence on modeling and tuning based on field-specific domain knowledge is low.


- 39 -

REFERENCES(1) K. Yano, “Invisible Hand of Data: The Rule for People,



(2) Hitachi News Release, “Hitachi Launches Hitachi AI

Technology/Business Improvement Service that Supports

to Resolve Corporate Management Issues through Artificial

Intelligence,” (Oct. 2015), http://www.hitachi.com/New/

cnews/month/2015/10/151026a.html.

(3) F. Kudo, T. Akitomi, and N. Moriwaki, “An Artificial

Intelligence Computer System for Analysis of Social-

Infrastructure Data,” Proc. of the 17th Conference on

Business Informatics (CBI), Vol. 1, IEEE (2015)

In order to strengthen intelligence technology

even further, Hitachi will continue working to make

it even easier to connect to existing systems, develop

new feature quantities, apply predictive diagnostics,

construct a service framework to enable application

to even more industries, horizontally deploy usage

and application logic with the assistance of the open

development community, and engage in other efforts

to further expand H.

Tomoaki Akitomi

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi Ltd. He is currently engaged in the research and development of AI and human behavior modeling. Mr. Akitomi is a member of the Japanese Society for Artificial Intelligence (JSAI).

Ryuji Mine

Center for Exploratory Research, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of AI, evolution science, learning science, and technology. Mr. Mine is a member of the IEICE, the JSAI, and the Information Processing Society of Japan (IPSJ).


Research & Development Group, Hitachi, Ltd. He is currently engaged in research and development work including AI in his role as Corporate Chief Scientist. Dr. Yano is a Fellow of the IEEE and member of the IEICE, The Japan Society of Applied Physics (JSAP), The Physical Society of Japan (JPS), and the JSAI. Author of “Invisible Hand of Data,” Soshisha.


Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of human information systems and AI. Dr. Moriwaki is a member of the Institute of Electronics, Information and Communication Engineers (IEICE), The Japan Society for Management Information (JASMIN), and the Association for Information Systems (AIS).

Fumiya Kudo

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi Ltd. He is currently engaged in the research and development of AI and big data analysis based on statistical approaches. Mr. Kudo is a member of the JSAI.

Toshio Moriya, Ph.D.

Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi, Ltd. He is currently engaged in the R&D management of AI and robotics, computer vision, and big data applications. Dr. Moriya is a member of the IEICE and the IEEE.

ABOUT THE AUTHORS


- 40 -

Featured Articles

AI Services and Platforms

A Practical Approach to Increasing Business Sophistication

Yasuharu Namba, Dr. Eng.

Jun Yoshida

Kazuaki Tokunaga

Takuya Haraguchi

OVERVIEW: Companies have begun working on business improvement initiatives inspired by insights acquired from events occurring around the world. To provide efficient support for these initiatives, Hitachi is working on projects to help solve wide scale problems and assist business growth. Collectively referred to as Hitachi AI Technology, these projects consist of cutting-edge AI technology and solutions driven by AI technology. The Hitachi AI Technology/Business Improvement Service was developed as the first of these projects. The service is designed to help solve management problems in areas such as improving corporate sales and cutting costs. This article describes an overview of this practical AI-driven service designed to increase business sophistication, and looks at the platforms that help make this service possible.

INTRODUCTION

TODAY’S more demanding worldwide market competition is making it difficult just to maintain the same prices for existing standard products and services. This environment is shifting the source of corporate competition toward innovation, that is, the creation of new value. Generally, value creation most often starts from a rediscovery or insight, but people tend to cling to fixed ideas formed from experience. They are often unable to free themselves from a limited range of ideas, and this tendency becomes more pronounced with longer experience. However, at the same time, the difficulty of defining goals (outcomes), recognizing the significance of business data, spotting exceptions, and applying value discoveries to business processes makes value creation difficult for anyone other than highly experienced insiders who are thoroughly familiar with the business. In response to these challenges, Hitachi decided to draw on the strengths of the latest artificial intelligence (AI) technology to boost human understanding and abilities, while creating new value to help solve wide scale problems and assist business growth through human-AI cooperation.

The recent rapid growth and spread of technologies such as cloud computing, mobile terminals, social media, and sensor technology is increasing the amount

of data being generated worldwide. Companies have started to draw on these technologies to gain an understanding of various events through data, to learn new insights from this data, and to apply them to policies leading to innovative business improvements. AI is an increasingly promising technology for efficiently assisting these efforts. For example, among sites that have nearly reached the limit of possible improvement with current business methods, there is a lot of demand for AI that can check whether human-created hypotheses are correct or not, or for AI that can devise hypotheses beyond human capabilities. Table 1 shows examples of demands for the use of AI to increase business sophistication in the area of marketing.

Industry/innovation demands

• Provide services tailored to the interests and preferences of individuals

• Improve product stock forecasting precision• Optimize overall expenses such as labor costs and

capital investment costs

Expectations for AI

• Identify elements that transform customer purchasing behavior

• Identify differences between customers of brick-and-mortar stores and online stores

• Identify characteristics of efficient sales activities

AI: artificial intelligence

TABLE 1. Examples of Demand for Use of AI to Increase Marketing Business SophisticationThere is increasing demand to identify trends that have been difficult for humans to notice previously, to provide services efficiently, and to innovate business.


- 41 -

To meet these expectations, Hitachi is working on projects to help solve wide scale problems and assist business growth. Collectively referred to as Hitachi AI Technology, these projects consist of cutting-edge AI technology and solutions driven by AI technology. The Hitachi AI Technology/Business Improvement Service(1) was developed as the first of these projects. The service is designed to help solve management problems in areas such as improving corporate sales and cutting costs. The following sections of this article describe an overview of the service, an example of its application to the creation of marketing solutions, and the Pentaho software used as the platform that helps make this service possible.

BUSINESS IMPROVEMENT SERVICE OVERVIEW

Using an AI technology developed by Hitachi called Hitachi AI Technology/H (hereafter referred to as H)(2), (3), the Hitachi AI Technology/Business Improvement Service creates business improvement proposals to help solve business problems. H is an AI technology that uses a large volume of complex business-related data to efficiently derive elements with strong correlations to organizational outcomes [key performance indicators (KPI)] and hypotheses for policies to improve them. The service uses H to propose improvement processes for problems facing various industries.

Expectations for AIUp until recently, experts who manage areas such as quality, sales, and stock have been studying policies for improving KPIs. However, with policy studies done by humans, there have been difficulties making objective evaluations because of assumptions based on preconceptions, stereotypical thinking, and personal hunches and experience. H is expected to overcome these shortcomings by eliminating preconceptions, discovering quantitatively important elements from data previously unused in analyses or proposed hypotheses, and proposing innovative improvement policies that do not rely on the thinking of human experts. But H is simply a tool, and the results obtained from it will vary greatly depending on which business processes its findings are applied to, and how they are used. So the best results are obtained by combining H with additional support services provided by technicians with the expertise to make thorough use of analytical methods and H.

Issues When Applying H to BusinessThe measures proposed by H can be appealing, often containing unprecedented suggestions. However, when applying them to actual business, the same sorts of issues faced by every company often have to be overcome (see Table 2).

APPLICATION OF AI TO BUSINESS IMPROVEMENT SERVICE

Marketing SolutionsHitachi’s Business Improvement Service uses AI to propose business improvements for areas such as retail sales, equipment maintenance, finance, and manufacturing. This section provides an example of how the service has provided marketing solutions for areas such as retail sales.

Conventional services analyze data collected from marketing systems [such as customer relationship management (CRM) systems and sales force automation (SFA) systems] to present suggestions to marketing staff with expert knowledge of marketing. However, conventional services are unable to incorporate marketing expertise and business site restrictions into the analysis.

Hitachi’s marketing solutions enable input and analysis of a variety of data such as business data sets, marketing staff expertise, site restrictions, and external environmental factors. Expertise that had previously only been tacitly understood as well as new insights can be turned into formalized knowledge and shared to increase the efficiency of the business improvement cycle (see Fig. 1).

In Hitachi’s marketing solutions, processes ranging from identifying current issues to measuring the benefits of marketing policies, making evaluations,

Application phase Challenges when applying

Utilizing for business

Even if new suggestions can be acquired by Hitachi AI Technology/H, it is unclear how to use them in business.

AI usage frequencyIt is unclear whether AI should be used once, or used repeatedly such as on a daily or monthly basis.

Selecting data parameters

New correlations may be found by increasing the number of data types or volume. But it is unclear how much more data is needed.

Preprocessing of data analysis

Before using AI, outliers must be removed from the data (data cleansing). If data contains many outliers, the analysis results can often be affected.

TABLE 2. Challenges when Applying AI to BusinessBelow are some examples of commonly encountered challenges when applying AI to business.

120 A Practical Approach to Increasing Business Sophistication

- 42 -

and proposing subsequent policies [the plan, do, check, act (PDCA) cycle] are categorized into 10 tasks. These tasks are combined in accordance with the client’s business conditions (see Fig. 2). Using H for Task 2 (Proposing improvement measures) enables comprehensive analysis that eliminates fixed ideas, leading to discoveries of new performance indicators for improving business issues (outcomes).

Using H to Formulate Improvement Measure ProposalsThis section looks at the example of using of H for retail sales marketing, with sales volume as the outcome. When conventional methods are used, the marketer decides on measures to improve sales using performance indicators derived from previous experience.

Customer attributes

Mar

ketin

g sy

stem

s (C

RM

, SFA

, etc

.)

Challenges

Scope covered by Hitachi’s solutions (creation of business practices for improvement cycle)

Scope covered by conventional analysis services (up to presenting suggestions to management department based on existing data)

Creating strategy proposals based on performance data and experiential

knowledge

Marketers

Implementing policies using instructions from management department and

impressions from site

Challenges

Shops Seasons

Weather

Lifestyles

Interests/preferences

Values

Customers

Implementing measures Purchasing

• Expertise tends to be personalized• Options quickly run out• Difficult to foresee benefits/find

corroboration

• Difficult to identify market changes (due to sudden trends, etc.)

• After identifying trends, it takes time until measures can be implemented

• Sudden responses are not possible

Data set Management department Business sites Changes in external

environment

Product information

Shop information

Purchase information (ID-POS)

Fig. 1—Scope of Marketing Solution Services.Marketing solution services provide business improvement cycles including business practices done before/after marketing systems.

ID: identification POS: point of sale CRM: customer relationship management SFA: sales force automation

Phase

Work on business

improvements

Benefits

(1) Planning (2) Implementation (3) Evaluation (4) Improvement

Identifying current state(such as KPIs and bottlenecks)

• Understanding current performance bottlenecks

• Understanding degree of impact of bottlenecks

• Ability to reliably and efficiently implement measures that meet objectives (stable operation)

• Ability to quantitatively measure improvement benefits

• Ability to quantitatively evaluate impact on performance based on measurement results

• Accumulate expertise about success/failure patterns in the form of performance

• Ability to save labor through automation of success patterns

• Ability to set tasks using quantitatively identified numerical values

• Evaluation indicators become clear, understanding measurement feasibility

• Understanding implementation/evaluation preparations and schedules, man-hours (costs)

1Proposing improvement measures

2Implement improvement measures

5

Evaluate KPI improvement benefits made possible by improvements

6

Examine successes and failures (incorporate failures into improvement plans)

8

Create regular business practices from measures (automation)

9

Automate measurement/evaluation

10

Evaluate impacts on performance

7

Designing evaluations/investigations of improvement benefits

3

Proposing implementation plans for improvement measures

4

Fig. 2—Tasks for Providing Marketing Solutions.These tasks provide comprehensive support for planning business innovation, implementing measures, evaluating benefits, and making improvements.



- 43 -

However, analysis using H involves a comprehensive

search for measures that improve outcomes, enabling

the discovery of previously overlooked effective

indicators and important indicators that tend to be

missed. Expertise previously considered to be the tacit

knowledge of the marketer is turned into formalized

knowledge from data, enabling the derivation of

new performance indicators. These indicators would

previously have been considered to be the marketer’s

hunches, but since they can now be corroborated by

data, they are expected to provide backing for new

initiatives (see Fig. 3).

PLATFORM SUPPORTING BUSINESS IMPROVEMENT SERVICE: PENTAHO SOFTWARE

Pentaho SoftwarePentaho software(4) is a data integration and analysis

platform used to integrate a wide variety of data created

from sources such as business systems, sensors,

and social media, and to analyze it from various

perspectives. Two platforms provide the environment

needed for all operations ranging from data collection

to analysis/usage. Pentaho Data Integration (PDI)

collects, processes, and outputs data, while Pentaho

Business Analytics (PBA) analyzes the collected data

and provides visual representations of it.

Pentaho software offers benefits that are not

available in competitors’ products. For example, it

enables data integration and analysis to be done on a

single platform, shortening the data usage cycle. It also

provides an abundant array of connected parts. And

since it is an open source software (OSS) product, it

can be quickly adapted to big data technology.

Using Pentaho Software in the Business Improvement ServicePentaho software is positioned as a data usage platform,

where data integration is performed by PDI, the data is

then analyzed by H, and finally PBA provides visual

representations of the results (see Fig. 4).

Data integration consists of creating visual

representations of the data provided by the client to

identify the data distribution and attributes (profiling),

remove heterogenous data from the original data

(cleansing), and join the remaining data to create a

data set. These processes are the preprocessing done

to enable analysis, and account for over half of the

entire analytical work. This preprocessing must be

done carefully since it can affect the analysis results if

OutcomeKPI candidates correlated

with outcomes

Customer axisWho contributes to sales?

Product axisWhat products contribute

to sales?

Conventional method(data + experiential knowledge)

Setting direction of measures using indicators that are already set

Experiential knowledge of marketers

Difficult to improve indicators/add new indicators

Purchase history

Sales promotion performance

Forecasting from experience

Insights

• Flexible creation of new indicators• Discover important indicators that

humans overlook

• Create formalized knowledge from personal expertise by analysis of sales promotion history and application to KPIs

• Corroborate significance of indicators using data

• Expertise tends to be personal• Options quickly run out• Difficult to foresee benefits/find

corroboration

When using AI(+ Hitachi AI Technology/H)

Sales volume

Fig. 3—Application of AI to Proposal of Improvement Measures and Expected Benefits.The use of AI to discover new indicators enables efficient proposal of innovative policies.

Pentaho DataIntegration

Pentaho Pentaho

PentahoBusinessAnalytics

HitachiAI Technology/H

Artificial intelligence

Cleanses and joins data provided by client to create a

data set

Inputs data set into AI and

analyzes

Compares data provided by client with analysis

results, and creates visual representations from a variety of perspectives

25%

Fig. 4—Link Between H and Pentaho Software.Data is integrated by Pentaho Data Integration, analyzed by H, and then visually represented by Pentaho Business Analytics.

122 A Practical Approach to Increasing Business Sophistication

- 44 -

on collaborative innovation activities with clients and

partners, promoting projects that help solve societal

issues and aid business growth by applying AI in a

wide range of areas.

REFERENCES(1) Hitachi, Ltd., Hitachi AI Technology/Business Improvement

Service, http://www.hitachi.co.jp/products/it/bigdata/

approach/ai-analysis/ in Japanese.

(2) K. Yano, “AI for Taking on the Challenges of an Unpredictable

Era,” Hitachi Review 65, pp. 14–34 (Jul. 2016).

(3) N. Moriwaki et al., “AI Technology: Achieving General-

Purpose AI that Can Learn and Make Decisions for Itself,”

Hitachi Review 65, pp. 35–39 (Jul. 2016).

(4) Hitachi, Ltd., Pentaho Software, http://www.hitachi.co.jp/

products/it/bigdata/platform/pentaho/ in Japanese.

done inadequately. The profiling process is done using

tools such as the R programming language.

The development environment provided by PDI is

provided by means of a graphical user interface (GUI),

and most operations can be done without programming.

PDI jobs (a job is a series of operations that are grouped

together) used in past projects can be modularized,

enabling them to be used immediately in new projects

just by making changes in the applicable locations.

Modularization also increases productivity. Even users

with no programming experience can combine and

define jobs using the GUI, enabling data processing

to be done easily. Java* coding will be needed if the

required data integration processes cannot be achieved

using only the standard connection and processing

parts provided by PDI. However, an abundant array

of Java methods specialized for data processing are

provided, enabling flexible and efficient processing.

Performance data shows that one project involving

Java data processing required 16.2 man-days for

cleansing and data integration, but only 7.5 man-days

for the same processes to be done using PDI, a labor

reduction of about 54%. When modularized templates

were applied to the same project, the time was further

reduced to 3.0 man-days, a labor reduction of about

81% relative to the original Java data processing.

PBA is used to load and create visual representations

of the H analysis results and data provided by the

client, evaluating it from a variety of perspectives.

The benefits of increases in the data volume on

performance are relatively small.

In the future, using the Pentaho software as

the data processing platform of the Hitachi AI

Technology/Business Improvement Service, Hitachi

will create templates on it for a wide variety of use

cases, aiming to further increase the efficiency of the

data integration, analysis, visual representation, and

evaluation processes to shorten the process cycle time.

CONCLUSIONS

This article has described a marketing solution that is

one of the solutions provided by Hitachi’s Business

Improvement Service, and the Pentaho software,

which is the platform technology that supports it.

The Hitachi Group is studying various types of

AI-driven business initiatives, through Group-wide

collective efforts. While making use of the successes

it has achieved to date, Hitachi will continue working

*Java is a registered trademark of Oracle and/or its affiliates.

Takuya Haraguchi

Big Data Solution and Services, IoT & Cloud Services Business Division, Service Platform Business Division Group, Information and Communication Technology Business Division, Hitachi, Ltd. He is currently engaged in the big data business.

Yasuharu Namba, Dr. Eng.

AI Business Development, Digital Solution Business Development, Service Platform Business Division Group, Information and Communication Technology Business Division, Hitachi, Ltd. He is currently engaged in the research and development of services utilizing AI. Dr. Namba is a member of the IEEE, the Association for Computing Machinery (ACM), the Information Processing Society of Japan, The Japanese Society for Artificial Intelligence, and the Society for Serviceology.

Jun Yoshida

Big Data Business Development Center, Digital Solution Business Development, Service Platform Business Division Group, Information and Communication Technology Business Division, Hitachi, Ltd. He is currently engaged in consultation for data analysis services.

Kazuaki Tokunaga

AI Business Development, Digital Solution Business Development, Service Platform Business Division Group, Information and Communication Technology Business Division, Hitachi, Ltd. He is currently engaged in the research and development of services utilizing AI.

ABOUT THE AUTHORS


- 45 -

Featured Articles

Utilization of AI in the Financial Sector

Case Study and Outlook for FinTech Era

Kiyoshi Kumagai

Satomi Tsuji

Hisanaga Omori

OVERVIEW: A succession of new user-oriented services combining finance and IT have been appearing recently. Known as FinTech, these services influence competitive advantage for their ability to create innovation beyond the framework of conventional financial services. Acting as an enabler for creating innovation, Hitachi is taking a close look at AI, and working on its utilization in the financial sector. This article looks at phenomena such as FinTech and the IoT as precursors of changes in the financial industry. It discusses how AI is being used in the financial industry through activities designed to raise management KPIs by quantitatively evaluating organizational activity levels from action data acquired from wearable terminals. The article also describes the future outlook for AI use by examining concepts such as the development of business applications that use embedded AI.

INTRODUCTION

FINANCIAL services have developed along with

advances in information technology (IT). The use of

IT among financial institutions extends to every aspect

of their business, and IT has become an indispensable

part of finance. However, until recently, IT has mainly

been used to improve the efficiency of operations

performed by humans.

New services called FinTech, due to their

combination of finance and IT, have recently been

gaining attention, and financial services offering users

a high level of convenience have appeared one after

another(1). These services go beyond the framework

of existing financial services, creating new value for

clients. The financial industry is on the verge of a new

era of innovation that will radically transform business

models through IT.

Acting as an enabler for creating innovation,

Hitachi is taking a close look at artificial intelligence

(AI) and working on introducing it to the financial

sector. In particular, anticipating the arrival of the

Internet of Things (IoT) era, Hitachi is taking on

the challenge of new services that support business

optimization that use AI to analyze sensor data from

wearable terminals, etc. that was unobtainable by

financial businesses in the past.

This article looks at how FinTech and the IoT

are changing the financial business environment,

and describes how AI is being used in the financial

sector through activities designed to improve the

management key performance indicators (KPIs) of

financial institutions by using AI to analyze action

data obtained from wearable terminals. The article also

describes the outlook for the use of AI in the financial

sector by looking at concepts such as the development

of business applications that use embedded AI.

CHANGES IN THE FINANCIAL BUSINESS ENVIRONMENT

Appearance of FinTechFinTech is a portmanteau word coined from Finance

and Technology. It is characterized by the creation

of innovative financial services that offer users a

high level of convenience by combining finance and

IT. The succession of FinTech services that have

appeared include user-to-user fund transfer services

the use Internet peer-to-peer (P2P) communication

technology, and cloud funding services that enable

direct procurement of funds from multiple individuals

through social media. These services have gained a

large amount of user support for their low cost and

procedural simplicity.

124 Case Study and Outlook for FinTech Era

- 46 -

New FinTech services are being provided by

IT industry startup companies, etc., which were

previously not major names in the financial industry.

As a result, the ability to create innovation that can

drastically transform financial business models using

technology is influencing competitive advantage.

AI has gained attention as the core technology

of FinTech. Financial services driven by AI include

new credit services and investment support services

that improve credit/risk management precision by

using AI to analyze Internet-based activity data and

transaction data. The use of AI is expected to produce

more advanced credit and risk management models

through discovery of new factors that were previously

undiscoverable.

Arrival of the IoT EraThe IoT is creating a network of objects that are

connected to the Internet, and its growth is driving the

innovation of services that use object operation data

and human activity data acquired from sensors. IoT-

driven innovation has become influential in fields such

as production management and product maintenance in

the manufacturing industry. Its use should continue to

advance in finance and many other areas in the future.

Several new IoT-driven insurance services are

appearing and growing in popularity. Examples include

telematics insurance in which driving data acquired

from vehicle onboard sensors is used to evaluate

accident risks to adjust insurance premiums, and new

medical insurance that evaluates the risk of illnesses

from health data acquired from wearable terminals.

Creation of Innovations in FinanceIn the IoT era, it will become possible to use sensors

attached to objects and people for realtime acquisition

of various types of data used in various applications.

The IoT era will see many innovative services appear

one after another. They will be made possible by

finding new connections by merging external data

acquired from sources such as the IoT, with internal

data gathered from financial operations.

Joseph Schumpeter, the father of innovation research,

said that innovation is created from new combinations

of existing things. For financial institutions to

constantly create innovation, it is important for them

actively gather external data from sources such as the

IoT, and to discover new connections by combining this

external data with their own internal data.

But discovering new connections from large

volumes of external data that change daily is not easy.

Hitachi is looking closely at AI as a technology that

exceeds human abilities to discover new connections,

and is working on developing analytical methods for

finding new correlations among massive volumes

of variables. The next chapter describes how AI is

being used in the financial sector through activities

designed to improve management KPIs by using AI to

analyze human behavior data obtained from wearable

terminals.

AI USE IN FINANCIAL SECTOR

Background and AimsTo respond to the changing financial business

environment, Hitachi is working on financial service

applications of the IoT and AI. One of these efforts

was a study on how to use technology for measuring

and analyzing human behavior(2) to improve the quality

of financial institution services and to help innovate

work styles. It included a trial conducted with The

Bank of Tokyo-Mitsubishi UFJ, Ltd. The trial studied

40 office workers from the planning department

of the bank’s headquarters. Differences in action

characteristics on days of high and low levels of

organizational activity(3) were extracted as proxies

for productivity indicator values, and it was found

that specific knowledge on productivity improvement

could be extracted.

Management Support Driven by Human Action Measurement and AIHitachi has developed technology for ID card-type

wearable sensors used to measure the actions of people

in groups. By connecting this technology to AI, it is

studying the creation of management support systems

that present quantitative advice on working styles that

increase productivity (see Fig. 1).

In a previous study, Hitachi acquired action data

about communication and deskwork from several

hundred subjects in one-second increments. However,

quantitatively expressing the actions and subjects

that contributed the most to the organizational

activity level was a costly and time-consuming

task requiring careful analysis by experts. The aim

of the recent study was to enter a large volume of

sensor data in the Hitachi AI Technology/H artificial

intelligence system (hereafter referred to as H) to

speed up comprehensive searches and the refinement

of objectives, and to enable analysts to focus on

presenting advice tailored to the client’s industry/

job type characteristics.


- 47 -

Trial Procedure and Analytical ResultsAction data was collected for three weeks at The

Bank of Tokyo-Mitsubishi UFJ, Ltd. and, using the

procedure shown by items (1) to (3) in Fig. 1, a search

was conducted for correlated action indicators, with

organization activity level as the target variable.

Items (4) to (6) were done by analysts, since work

characteristics needed to be taken into account.

To quantify typical office worker work patterns,

groups of indicators related to deskwork and

communication were used as action indicators. The

indicators that were collected daily were entered into

H (see Table 1). H generated compound indicators

combining attributes and actions, refined them into

effective indicators that could describe organizational

activity levels, and generated stochastic models.

Fig. 2 shows some of the results. The first finding

was that days on which subjects in their 30s had

short, frequent conversations had higher overall

organizational activity than other days. A difference

in action indicators between new and experienced

department members was also found. The activity

level was better when new department members (with

less than 3.5 years in the department) did prolonged

deskwork (at least 30 minutes). However, for more

experienced department members (with at least 3.5

years in the department), the activity level increased

more when deskwork was divided into sessions of

less than 30 minutes due to interruptions for reasons

such as answering questions. This finding shows

that although the personal productivity of more

experienced workers may decline, they improve

the productivity of the overall organization. This

corroborates the notion that more experienced workers

contribute to raising team synergy. Quantitatively

expressing this contribution to the entire organization

Level of activity in

organization

Individual action

indicators

Action data

(1) Action measurement

ID card typewearable sensors

(6) Changes to actions and work processes

(2) Compound indicator generation

(3) Generation of stochastic model, sorting of valid indicators

(4) Application of industry/job type characteristics

(5) Presentation of advice statement

Server Hitachi AITechnology/H

Physicalworld

Application for specific industry/job type

Fig. 1—Conceptual Diagram of a Management Support System.AI is used to analyze collected human behavior data, to provide advice on work methods designed to increase productivity.

Category Indicator Definition

Duration of interaction (min)

Total time Time during which interaction with at least one other person is detected

Two-way Time during which two-way conversation is in progress

Pitcher Time during which subject is speaking

Catcher Time during which subject is listening

Number of interactions: Number of instances of each category of interaction duration

(a) Continuing for < 5 minNumber of short conversations (greetings or passing on a message)

(b) Continuing for 5 < 15 min

(c) Continuing for 15 < 30 minNumber of long conversations (such as meetings)

(d) Continuing for ≥ 30 min

Duration of deskwork (min)

Total duration of deskwork Time during which subject does not interact with others and has minimal physical movement

Maximum duration of continuous deskwork

Longest period of uninterrupted deskwork during the day

Number of instances of deskwork: Number of instances of each duration category

(a) Continuing for < 5 min Number of times deskwork is interrupted (by being spoken to, goingfor a walk, etc.)(b) Continuing for 5 < 15 min

(c) Continuing for 15 < 30 minNumber of times deskwork continues for a long period with few interruptions

(d) Continuing for ≥ 30 min

Length of time sensor is worn (min)

Length of time sensor is wornTime measured by ID card type sensor (in the case of office work, this is the office’s working hours)

TABLE 1. List of Action Indicators for Office Workers

Hitachi has defined indicators for interactive communication and deskwork.

126 Case Study and Outlook for FinTech Era

- 48 -

may help experienced workers understand the value

of their support to others, and be effective for overall

optimization of service and work methods.

Validity and Remaining ChallengesThe case study described in the previous section

confirmed the validity of the following:

(1) It was possible to extract knowledge for improving

service and work methods in a financial institution

through quantitative analysis of data from wearable

terminals.

(2) It was possible to use AI to comprehensively

scrutinize which employees to focus on, and which

actions of those employees to focus on.

(3) It was possible to use AI to extract individual

indicators for overall optimization.

Achieving ongoing operation by automating

the presentation of advice is a remaining challenge

for achieving a management support system. The

analysis of the results output by H are currently being

interpreted and compiled by analysts, and client work

characteristics are being considered when presenting

advice. To increase the efficiency of these processes,

it may be necessary to embed functions such as work

applications into H to automate the analysis, and to

create a method of visualizing and sharing results on

a daily basis.

FUTURE OUTLOOK

This section describes the future outlook for the use of

AI in the financial sector, looking at the projects that

are currently underway.

Creation of Social Innovation in the Financial SectorHitachi is focusing on it Social Innovation Business,

which solve problems in public systems by combining

infrastructure technology and IT. By using IT to link

public infrastructure projects in areas such as finance

and railroads with related peripheral projects, it aims to

produce innovative services by creating new connections

that exceed the framework of existing projects.

Expectations for innovation are increasing in the

financial sector, and AI has gained attention as an

enabler for the discovery of new connections. For

example, new FinTech services are being created

by linking finance and electronic commerce

(eCommerce), with functions such as calculation of

credit scores of eCommerce providers by using AI to

analyze the data of transactions on eCommerce sites.

Hitachi is working on optimization problem

searches in financial operations, an area in which

AI is effective. It seeks to speed up the creation of

Social Innovation in financial operations by using AI

to develop optimization models.

Combining Financial Operation Data and IoT DataThe arrival of the IoT era will greatly increase the scope

of data that financial institutions can use. Acquiring

position information and operation data from sensors

attached to objects such as vehicles and residential

facilities is already possible, and it is becoming

possible to acquire activity data and health data in

real time from wearable terminals worn by people.

The use of this data holds the potential to radically

transform the business models of financial institutions.

Specifically, the insurance industry is expected to

evolve from calculating risks using traditional statistics

to calculating risks for each policy individually using

the IoT.

Action (low)

Mea

n le

vel o

f ac

tivity

in o

rgan

izat

ion

30s/number of interactions lasting less than 5 min

9.7

0

5

10

15

Action (high)

12.5

Action(low)

Mea

n le

vel o

f ac

tivity

in o

rgan

izat

ion

Less than 3.5 years in department/30 min or longer of uninterrupted deskwork

8.9

0

5

10

15

Action(high)

11.4

Influence + 2.5

Influence + 2.8

Influence – 1.6

Action(low)

Mea

n le

vel o

f ac

tivity

in o

rgan

izat

ion

3.5 years or more in department/30 min or longer of uninterrupted deskwork

10.8

0

5

10

15

Action(high)

9.2

(a) Relationship between number of short conversationsby people in their 30s and level of activity in organization

(b) Different action guidelines depending onhow long employee has been in the department

Fig. 2—Example of Action Characteristics Analysis for Different Attributes*.(a) The level of activity in the organization is higher by 2.8 on days on which people in their 30s had frequent short conversations such as greetings or passing on a message. (b) Opposite guidelines were obtained for uninterrupted deskwork time depending on how long the employee has been in the department.

*The experimental data shown in this article (see Fig. 2) is a mock up for

presentation purposes. However, the knowledge obtained is the same as

that achieved in practice.


- 49 -

CONCLUSIONS

This article has discussed the expanding use of AI in the financial sector by looking at trends in the creation of innovations in finance in light of the changing financial business environment resulting from FinTech and the IoT. Also described were a study undertaken by Hitachi as the first step toward achieving these innovations in which organizational activity levels were measured using ID card type wearable sensors and AI, and the future outlook for AI.

In the future, the growth of the IoT should create an era in which the degree of skill in the use of external data will affect the competitive advantage of financial businesses. To prepare for this coming era, Hitachi wants to help create financial sector innovations by creating analytical methods that combine internal data from financial operations with external data, and working on developing embedded AI business applications.

REFERENCES(1) Nikkei Computer, ed., “The FinTech Revolution,” Nikkei

Business Publications, Inc. (Dec. 2015) in Japanese.(2) M. Hayakawa, N. Ohkubo, and Y. Wakisaka, “Business

Microscope: Practical Human Dynamics Acquisition System,” The Transactions of The Institute of Electronics, Information and Communication Engineers, Vol. J96-D, No. 10 (Oct. 2013) in Japanese.

(3) K. Yano et al., “Measuring Happiness Using Wearable Technology—Technology for Boosting Productivity in Knowledge Work and Service Businesses—,” Hitachi Review 64, pp. 517–524 (Nov. 2015).

(4) K. Yano, “Invisible Hand of Data: The Rule for People, Organizations, and Society Uncovered by Wearable Sensors,” Soshisha Publishing Co., Ltd. (Jul. 2014) in Japanese.

Creating analysis environments and methods by accurately combining in-house data acquired using existing financial operations with new external data acquired from the IoT will be an indispensable requirement for achieving these new business models.

To meet this requirement, Hitachi is creating AI analysis models that combine financial operation data and IoT data, and developing and applying methods of creating new financial services from the connections discovered by these models.

Development of Embedded AI Business ApplicationsTo maximize the benefits of AI-driven analysis, AI should ideally be embedded in routine operations, and new models for financial operations should be constructed that coordinate humans and AI.

To achieve these aims, there is a need for embedded AI business applications that provide support for routine decision-making using financial business applications with embedded AI. For example, an embedded AI application that proposes policy plans could be developed for the insurance industry that would enable optimum plan proposals to be derived by coordinating the efforts of AI and humans. By using AI to perform analysis that combines existing policyholder data with IoT-based action data and health data, this application could be expected to propose appropriate policy plans that capture the risks and preferences of individuals in real time.

In the future, Hitachi would like to develop embedded AI business applications that draw on the strengths of AI to bring new innovations to financial operations.

Satomi Tsuji

Global Center for Social Innovation – Tokyo, Research & Development Group, Hitachi Ltd. She is currently engaged in research into the use of human big data for organizational management. Ms. Tsuji is a member of The Society of Project Management.

Kiyoshi Kumagai

Global Center for Social Innovation – Tokyo, Research & Development Group, Hitachi Ltd. He is currently engaged in research into service business design methodology and application. Mr. Kumagai is a member of the Information Processing Society of Japan (IPSJ).

Hisanaga Omori

Financial Information Systems Sales Management Division, Financial Institutions Business Unit, Hitachi Ltd. He is currently engaged in developing solutions for financial institutions.

ABOUT THE AUTHORS


- 50 -

Featured Articles

Utilization of AI in the Railway Sector

Case Study of Energy Efficiency in Railway Operations

Ryo Furutani

Fumiya Kudo


OVERVIEW: Leveraging its past track record in rolling stock maintenance of high-speed Class 395 rolling stock for the UK’s High Speed 1 project, for which it received orders in 2005, Hitachi is advancing the expansion of its services business in the railway sector, for example, through the UK’s Intercity Express Programme (IEP) and by providing rolling stock maintenance for Abellio, a railway operator. Condition monitoring systems that remotely monitor the condition of rolling stock will be the key to expanding and developing the services business. Hitachi is working on utilizing AI technology that it has developed to provide further added value using rolling stock information that is collected on a daily basis. This article covers power consumption while operating railway rolling stock, and presents an analytical case study of where feature values for reducing power consumption are identified using AI.

INTRODUCTION

INCREASED energy efficiency of railway systems, both inside and outside Japan, is being sought for the purpose of reducing CO2 emissions as a measure against global warming(1). Sixty to eighty percent of the energy consumed by railway systems is the energy used when operating rolling stock, and increasing the energy efficiency is effective in reducing CO2 emissions. For this reason, Hitachi developed the A-train concept featuring a lightweight aluminum structure and main converters that apply silicon carbide (SiC) hybrid modules to achieve increased energy efficiency in the overall traction power supply system(2).

The utilization of data obtained by measuring operating rolling stock is pointed out as one means of verifying the energy savings from applying these technologies. With information and communication technology (ICT) progressing rapidly in recent years, there has been accelerated movement toward utilizing the diverse sensor information that is collected by railway systems in operation and maintenance (O&M) services. Hitachi, too, is expanding its rolling stock maintenance services through remote condition monitoring.

The use and application of artificial intelligence (AI) such as deep learning is being vigorously

promoted as a technique for high speed and efficient processing of the vast amounts of information collected by these technologies.

This article describes a case study of the application of Hitachi AI Technology/H (hereafter referred to as H) to the analysis of energy saving performance in terms of rolling stock energy, and the future outlook for railway systems where AI is put to use.

UTILIZATION OF AI IN THE ANALYSIS OF ENERGY SAVING PERFORMANCE IN TRACTION POWER CONSUMPTION

ApplicationsIn this case, some rolling stock operating data collected by remotely monitoring the condition of rolling stock was used to automatically extract the most effective feature values for reducing traction power consumption (i.e. the energy consumed by the traction power supply system when driving rolling stock motors) with H (see Fig. 1).

The technology on which H is based is a statistical technique in which the objective variables and explanatory variables must be assigned in advance. For this reason, the traction power consumption of the entire train per travel between stations at which the train stops as the objective variable for one sample, and the time-history data from the rolling stock operating


- 51 -

data at that time was directly assigned as explanatory variables (see Table 1).

Then, a distinctive technology of H, leap learning, was used to automatically and comprehensively generate the objective variables, correlations, and feature values having a large influence based on explanatory variables, including nine parameters of rolling stock operating data, such as the rolling stock travel speed (carriage speed), and three parameters of track infrastructure data, such as the track gradient information.

Furthermore, one year’s (2013) worth of data that was collected when a specific train passes through four stations was used as the input data for H.

Application ResultsApproximately 4,000 feature values were automatically generated by H based on the input data of the objective variable and explanatory variables shown in Table 1 (see Fig. 2).

The following shows one example of the feature values that were automatically generated: (1) Carriage speed is 0 to 57 km/h(2) Gradient is down gradient(3) Operating time is 18:00 to 24:00 and Mass of carriage A is 45,000 to 48,000 kg

These features can be broadly divided into three categories: “feature values (1)” that are directly generated from numerical data, “feature values (2)” that are directly generated from character codes, and “feature values (3)” that are combinations of individual

Fig. 1—Example of Utilizing H for Rolling Stock Operating Information.Diverse sensor information during rolling stock operation is input into H, and parameters for reducing traction power consumption are automatically extracted.

Traction power supply system

Motor Motor

Condition monitoring

system

Rolling stock information

Measures and hypotheses

HitachiAI Technology/H

No Item Item Type Unit

1 Objective variable Traction power consumption Number kWh

2

Explanatory variables

Rolling stock

operating information

Rolling stock travel speed Number km/h

3 Train mass Number kg

4 Individual carriage mass Number kg

5 External air temperature Number ˚C

6 Acceleration/deceleration Number m/s2

7 Coupling information Character -

8Operating

information (notch)

Character -

9 Up/down Character -

10 Operating date/time Character -

11

Track infrastructure information

Gradient Character -

12 Line feature value Character -

13 Curve information Character -

TABLE 1. List of Data Input into HTraction power consumption was provided as the objective variable, and rolling stock operating information and track infrastructure information were provided as explanatory variables.

Objective variable“Power consumption”

Feature values (approx. 4,000)Ex 1: Ex 2: Ex 3:

Carriage speed is 0 to 57 km/hGradient is down gradientOperating time is 18:00 to 24:00 and mass of carriage A is 45,000 to 48,000 kg

Fig. 2—Feature Values Generated by H.For H, approximately 4,000 feature values were automatically generated from the 12 explanatory variables.

130 Case Study of Energy Efficiency in Railway Operations

- 52 -

feature values. By analyzing the correlations between

the objective variables and these automatically and

comprehensively generated feature values, it is possible

to gain knowledge and hypotheses that humans cannot

process and that humans are not capable of noticing.

The following explains the effective feature

values that were extracted from the approximately

4,000 automatically generated feature values. The

following feature value, which has the highest

negative correlation (correlation coefficient: −0.81)

with traction power consumption, was extracted in

August 2013:

Feature value: Operating information (notch) is

Notch-off

The operating information feature value, notch,

expresses a step in acceleration force for accelerating/

decelerating rolling stock. For the railway rolling stock

discussed in this article, among the notch positions,

when the notch-off operating time is longer, traction

power consumption shows a downward trend (see

Fig. 3).

On the other hand, the extracted feature value

with the most positive correlation with traction

power consumption was the notch called “maximum

notch (correlation coefficient: 0.73).” This means the

maximum notch operating time should be lengthened.

The fact that these feature values, notch-off and

maximum notch, were extracted as effective feature

values indicates that both are largely affected by driver

operation. Also, in each of the other months of 2013,

it was confirmed that there was a high correlation

between each of the respective feature values and the

objective variable.

Moreover, the traction power consumption on the

vertical axis and feature value on the horizontal axis

in Fig. 3 have been normalized by the travel distance

of each of the four representative sections and by the

travel time of each sample, respectively.

Next, Fig. 4 shows the carriage speed information

and notch information for each of the following

operations, Operation 1 and Operation 2, in the sample

for representative Section A in Fig. 3.

Operation 1: Traction power consumption is large,

and notch-off operating time is short (August 26,

2013).

Operation 2: Traction power consumption is small,

and notch-off operating time is long (August 6, 2013).

According to Fig. 4, for the travel in Operation

1, travel under notch-off operation was conducted

frequently for short distances, and there were long

sections of travel under maximum notch operation.

Whereas, for the travel in Operation 2, travel under

notch-off operation was conducted infrequently and

over relatively long distances, and, in the latter half

of this section, it can be confirmed that the rolling

stock traveled in such a way that the section travelled

under notch-off operation was longer. From this fact,

it can be confirmed based on the data obtained by

measuring actual rolling stock operation, that there are

differences in driving skills even in the same section

of travel.

Estimating the Effect of Energy Consumption ReductionFig. 5 shows the correlation between traction power

consumption and notch operating times (notch-off,

maximum notch) in Section B for the period of one

year, 2013. According to the figure, the longer the

maximum notch operating time is, and the shorter

the notch-off operating time is, traction power

consumption increases. Alternatively, it can be

seen that a relationship exists where traction power

consumption decreases when the maximum notch

operating time is shorter and the notch-off operating

time is longer. Furthermore, it can be seen that there

is large variation in each of the samples and there is

room for improving traction power consumption.

In this respect, if we assume that operation has

improved in the 20% superset that has small traction

power consumption along the regression line in Fig. 5,

then a yearly decrease in traction power consumption

of approximately 20% can be anticipated. Furthermore,

the relationship between traction power consumption

1.6

1.4

1.2

1

0.8

0.6

0.40 0.2 0.4 0.6

Notch-off operating time (normalized)

Notch:Steps of force for accelerating/decelerating rolling stock

Tra

ctio

n po

wer

con

sum

ptio

n (n

orm

aliz

ed)

0.8

Operation 1

Operation 2

1

Section A

Section B

Section C

Section D

Fig. 3—Relationship between Traction Power Consumption and Notch-off Operating Time.There is an extremely high negative correlation (correlation coefficient: −0.81) between the notch-off operating time and traction power consumption of the four representative sections on August, 2013.


- 53 -

and notch operating time shown in Fig. 5 shows the

same tendency in the other three representative sections.

And, when the four representative sections are taken

as a whole, it can be confirmed that a yearly power

consumption reducing effect of approximately 14%

can be anticipated. Rolling stock operation information

in which rolling stock ran punctually according to

operating travel times was used for this analysis.

This case study introduced a study where H was

applied with data limited to a representative train

and four representative sections, however, Hitachi

is currently proceeding with analysis of expanded

travel distances using multiple trains. It is also

proceeding with analysis using an expanded amount

of information for explanatory variables such as the

operation status of the traction power supply system

which was not targeted as an explanatory variable in

this article.

Maximum service speed

Car

riag

e sp

eed

Car

riag

e sp

eed

Ope

ratin

g co

mm

and

(not

ch)

Ope

ratin

g co

mm

and

(not

ch)

Maximum notch

Distance

Distance

Notch-off

Notch-off travel section

Notch-off travel sectionMaximum service speed

Maximum notch

Notch-off

(a) Operation 1 (August 26, 2013)

(b) Operation 2 (August 6, 2013)

Fig. 4—Comparison of Carriage Speed Information and Operating Command Information in Section A.With travel under Operation 2 where the traction power consumption is small, travel was conducted under relatively long notch-off operations.

0.75

0.5

0.25

00 0.25 0.5 0.75

200

300

400

20% reduction

20% superset Regression line

Universal set

Notch-off operating time (normalized)

Max

imum

not

ch o

pera

ting

time

(nor

mal

ized

)

500(kWh)

Fig. 5—Visualization of Relationship between Notch Operating Time and Traction Power Consumption.For the year 2013, the variation in notch operating times was large, and when operation improvements were assumed, it was found that a power consumption reduction effect of about 20% could be anticipated for the year in representative Section B.

132 Case Study of Energy Efficiency in Railway Operations

- 54 -

be pointed out as representative KPIs in railway

maintenance. If wearable sensors are made use of

in maintenance services, measures for improving

and enhancing work efficiency can conceivably

be identified based on maintenance workers’ daily

activities. Moreover, maintenance workers’ level of

well-being (happiness) also could be applied as a

measure for improving work efficiency. By improving

the rolling stock utilization rate, it is anticipated that

the relationship between the time-related deterioration

of rolling stock and the operating conditions of rolling

stock will be discovered from H, and that this can

be applied to the detection of the warning signs of

equipment failure.

CONCLUSIONS

This article introduced a case study where H was

applied to automatically extracting feature values to

reduce the power consumed in driving rolling stock,

and described the future outlook for applications of H

in railway O&M services.

Hitachi intends to accelerate the full-scale

application of AI to the railway sector, and to promote

further initiatives for increasing energy efficiency in

railway operations and for improving efficiency in

rolling stock maintenance.

REFERENCES(1) Moving Towards Sustainable Mobility—A Strategy for 2030

and beyond for the European railway sector, UIC (2012).

(2) T. Mochida et al., “Development and Maintenance of Class

395 High-speed Train for UK High Speed 1,” Hitachi Review

59, pp. 39–46 (Apr. 2010).

FUTURE OUTLOOK OF RAILWAY SYSTEMS THAT UTILIZE AI

In the future, it is anticipated that H will be applied

in various situations on diverse big data that is

collected by railway systems. In particular, the applied

deployment of H in railway O&M services that is

being promoted by Hitachi is described below (see

Fig. 6).

In railway operations, comfort, etc. is one

representative key performance indicator (KPI) in

addition to the increased energy efficiency introduced in

the applied case study mentioned above. For increased

energy efficiency, it is conceivable that the power

consumption of the auxiliary power supply system

used, for example, for operating air conditioning or

opening/closing doors, will be targeted in addition to

the power consumption of the traction power supply

system. The extraction of new knowledge can be

anticipated since power consumption is affected more

substantially by the behavior of people in the carriages.

And since there are also two ways of operating, by

electric rolling stock and by rolling stock with diesel

engines, depending on carriage composition, the

identification of increased energy efficiency measures

in operation management can also be anticipated. With

regard to comfort, comfort parameters relating to ride

quality, such as vibration and noise can be targeted

and design guidelines for operating rolling stock

comfortably and safely may conceivably be gained.

Next, indices relating to the work efficiency of

maintenance workers and the rolling stock utilization

rate resulting from rolling stock malfunctions can

Increased energy efficiency• Power consumption(Traction power supply/auxiliary power supply system)

Safety/reliability• Rolling stock operating rate(Detect warning signs of equipment failures)

• Improved business efficiency(Simplifies work/reduces inventory)

Comfort• Ride quality (Noise/vibration)

• Happiness(Workplace environment)

Operation Maintenance

Fig. 6—Example of Applicable Targets on Railway O&M Services.This figure shows some applicable targets for H with respect to increased energy efficiency, comfort, and safety/reliability.


- 55 -

Fumiya Kudo

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of AI and big data analysis based on statistical approaches.

Ryo Furutani

Transportation Systems Research Department, Center for Technology Innovation – Mechanical Engineering, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of condition monitoring systems for railways. Mr. Furutani is a member of the The Japan Society of Mechanical Engineers (JSME).


Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of human information systems and AI. Dr. Moriwaki is a member of the Institute of Electronics, Information and Communication Engineers (IEICE), The Japan Society for Management Information (JASMIN), Association for Information Systems (AIS).

ABOUT THE AUTHORS


- 56 -

Featured Articles

Use of AI in the Logistics Sector

Case Study of Improving Productivity in Warehouse Work

Junichi Hirayama

Tomoaki Akitomi

Fumiya Kudo

Atsushi Miyamoto, Ph.D.

Ryuji Mine

OVERVIEW: Attempts are being made to increase the efficiency of work improvements through more widespread application of IT to work systems. However, as each new improvement is added or improvements are made with respect to environmental changes, it requires manual changes to the system, leading to increases in work improvement costs. Hitachi has developed an AI system that uses big data such as work performance information, to understand worksite improvements and environmental changes and issue appropriate work instructions. It has conducted a demonstration test, which confirmed the effectiveness of this system for improving distribution warehouse work. In the future, Hitachi will continue to work on expanding the AI system to a wide range of Social Innovation Business in areas such as manufacturing and distribution.

INTRODUCTION

AS information technology (IT) has advanced in recent years, IT systems have been introduced in a variety of work to improve efficiency. In the future it will be important to make further improvements to work efficiency that take site improvement activities and site environmental changes into consideration. However, conventional work systems are controlled by pre-designed programs, and system engineers have had to redesign the systems whenever new improvement activities needed to be applied to work systems. Moreover, work procedures and settings have had to be changed whenever changes in the work environment necessitated operations that were different from the current conditions. Such frequent system changes have been expensive and have made it difficult to issue efficient work instructions rapidly to respond to new improvement activities or environmental changes.

This article describes the development of an artificial intelligence (AI)-driven work system that uses big data from work performance information gathered on a daily basis by the work system, to issue appropriate work instructions by understanding worksite improvements or environmental changes. The AI-driven work system was subjected to demonstration

testing at a distribution warehouse to determine its effectiveness, and these test results are also described.

AI-DRIVEN WORK SYSTEM USING HITACHI AI TECHNOLOGY/H

Hitachi AI Technology/H (hereafter referred to as H) is an original AI system that the Research & Development Group at Hitachi, Ltd. developed as a means of achieving an AI-driven work system. This chapter describes the features of the AI-driven work system developed based on this technology.

Deriving Work Improvement Proposals Originating from DataH is an AI system used for data analysis to automatically calculate relationships between key performance indicators (KPIs) and the explanatory variables related to them. Specifically, it generates hundreds of thousands of feature values by comprehensively combining explanatory variables on an exploratory basis, and describes the relationships between the KPIs and these feature values in the form of equations. For example, taking site work efficiency as the KPI and work behaviors (people, locations, things, quantities, etc.) as the explanatory variables, it can generate


- 57 -

numerical models of behavioral characteristics related to work efficiency. These models can be used to derive proposed measures for work improvements that originate from the data (see Fig. 1).

Understanding Human Ingenuity and Refl ecting It in Work InstructionsFig. 2 shows the configuration of the AI-driven work system. Site workers perform work according to the work instructions output by the work system, and the results of the work done by the workers are collected daily by the work system. The AI-driven work system

uses the work results collected by H to derive work improvement proposals, and reflects them in the work instructions.

The site workers work according to the work instructions output by the work system, however, to work efficiently, they often add their own ingenuity or improvements to the work based on their own experience. H recaptures and analyzes the results created from worker ingenuity and improvements, to select the results that generate higher efficiency and reflects them in subsequent work instructions. By understanding site workers’ ingenuity and improvements and repeatedly reflecting them in work instructions on a daily basis, H makes it possible to continually improving work efficiency through mutual cooperation between humans and AI.

Rapidly Incorporating Various Types of Big DataThe big data collected by work systems consists of several different types of data such as numerical quantities, times, and product codes, along with text and symbols. So, for all this data to be entered into an analytical system, it must be tagged in advance by experts with knowledge of the industry and business operations, requiring work whenever data is added or changed. H has a function that rapidly enters new additional data without requiring human intervention. The function works by analyzing the statistical distribution of the data and automatically identifying data formats such as quantities, times and product codes in advance. This enables daily worker ingenuity and fluctuation in demand to be automatically entered into the system and reflected in the work instructions in a timely manner (see Fig. 3).

Automatically interprets input data, generates combinations

Input dataHundreds of thousands of feature values

Mea

sure

s

Equations:Outcome = f(x)

Refines feature values related to outcomes, and derives relationship equations

Fig. 1—Overview of H.H generates feature values from input data (explanatory variables) comprehensively combined on an exploratory basis and describes their relationships to the outcomes (KPIs) in the form of equations.

Work system Worksite

AI

Work big data (past performance

data)

Work request

Client

Derive work improvement proposals originating from data

Rapid input of big data in various formats

Understand ingenuity from work results and reflect them in work instructions

Work instructions including ingenuity derived from data analysis

• Add original ingenuity• Obtain work improvement hints from work instructions

Work results incorporating human ingenuity

Generate work instructions

Fig. 2—AI-driven Work System Configuration.The system enables continual improvement of work efficiency through human-AI cooperation by issuing AI-derived work instructions and a cycle of collecting work results incorporating human ingenuity.

136 Case Study of Improving Productivity in Warehouse Work

- 58 -

DEMONSTRATION TEST FOR DISTRIBUTION WAREHOUSE WORK

Challenges of Distribution Warehouse WorkAs the importance of logistics in the distribution industry increases, improving distribution warehouse work is becoming indispensable for maintaining competitiveness. Specifically, shorter work times are needed for work processes such as receiving and shipping. Receiving consists of receiving products from shippers and storing them at the designated locations in the warehouse. Shipping consists of receiving orders from stores or individuals and picking (collecting) products stored in the designated locations. The aim of this demonstration test was to reduce the work time spent on picking work, which has the highest work cost.

Test OverviewPicking work consists of collecting a specified product from the warehouse in response to a product order from a client. The picking work instructions specify the product to collect and the product’s storage location. The worker follows these instructions to

travel through the warehouse and collect the specified product. The warehouse management system (WMS) issues the picking work instructions and collects the work results.

Fig. 4 shows the configuration of the AI-driven work system used in the demonstration test. The data flow in the standard work system (WMS) is illustrated by (3) (4), while the data flow in the AI-driven work system is illustrated by (1) (2) (3) (4).(1) H reads past work results, and generates mathematical models of the KPIs and work behaviors.(2) The generated models are used to generate improvement proposals for that day’s work instructions. The work instruction improvement proposals are fed back to the WMS.(3) Work instructions are issued from the WMS.(4) The work is done as specified in the work instructions, and the work results are collected in the WMS.

In other words, the difference between the standard work system and the AI-driven work system in this test was whether standard work instructions are issued in Step (3), or whether the work instructions devised by H are issued.

(1) Mean, median and mode position relationships

(2) X-axis differential distribution

(5) Text string notation pattern(6) Edit distance and distribution

of category text strings

(4) Y-axis differential distribution

(3) Skew, kurtosis

Identify data types

Extract statistical distributions, notational variation and other characteristics

Numerical values, quantities, dates, times, (required) durations, orders, categories, IDs, item numbers, codes, names, etc.

Data A70345

7 0 3 4 5

7 0 3 4 87 0 2 3 27

count

Symbol Number (list) Number (list)Dot

ED = 1

ED = 3

=mean =val =

170

100

22.50.8

33.22.5

105.15.4

0 2 2 7

7034870232

70227S + E.[0-9] [0-9]

−

Fig. 3—Example of Automatic Identification of Data Types by H.H combines statistical distributions of data and notational knowledge to automatically identify data types.


- 59 -

Analysis ResultsThe mathematical model results [the results of Step (1)

in the previous section] generated by H are described

below. The picking work times were set as the KPI,

and the work attributes of picking work (what work

was done, how much was done, and by whom, when

and where) were set as the work behaviors.

Worker congestion at specific times and locations

(aisles) in the warehouse was obtained as a behavioral

characteristic that had a large effect on picking work

time. Fig. 5 is a graph showing the degree to which

each aisle’s work time was affected. As it shows, the

higher the value for an aisle, the more the work time

tended to increase when that aisle was crowded.

Using the mathematical models, H created the

day’s work instructions in a way designed to reduce

worker congestion at particular times and locations

[Step (2) in the previous section].

Demonstration Test ResultsHitachi conducted a trial of the AI-driven work

system over a period of about two months in an

actual distribution warehouse, and compared the

work efficiency to the work efficiency during a control

period. The trial was not explained to the site workers,

and they worked as usual according to the picking

work instructions.

Fig. 6 is a histogram of the picking work times.

The horizontal axis shows the work time, and the

vertical axis shows the number of work operations.

The histogram bars representing work operations done

during the trial period are shifted to the left relative

to the bars for the operations done during the control

period, indicating an overall reduction in work time.

Work time was reduced an average of 8%.

CONCLUSIONS

This article has described an AI-driven work

system that uses big data from work performance

information to issue appropriate work instructions

(2) Propose picking work instruction improvements

(3) Issue picking work instructions

(1) Load past work results(KPI + Work attributes)

(1) Load picking work instructions

(4) Collect work results

Shows the flow of data in a conventional work system.

Shows the additional flow of data in the AI-driven work system.

Work system(WMS)


Worksite(distribution warehouse)

Fig. 4—Configuration of AI-driven Work System Used in Test.The data flow of the conventional system is shown by (3) (4). The AI-driven work system adds the new data flow shown by (1) (2).

WMS: warehouse management system

Degree of effect on work time

Con

gest

ed w

ork

aisl

e

0

F

H

L

M

P

Q

S

T

U

X

Y

Z

AA

AB

AD

0.1 0.2 0.3 0.4 0.5

Fig. 5—Congestion and Its Effect on Work Time for Individual Work Aisles.This histogram shows that the congestion of specific work aisles has a large effect on work time. (The work aisle names are sorted in order from greatest to least effect on work time.)

Work time (seconds per product)

Picking time was reduced

Num

ber

of w

ork

oper

atio

ns

00

2

4

6

8

10

12

14

16

20 40 60 80 100 120

Control period

Trial period

140 160

Fig. 6—Histogram of Work Operations by Picking Work Time.An overall reduction in work time was demonstrated for the trial period relative to the control period.

138 Case Study of Improving Productivity in Warehouse Work

- 60 -

REFERENCES(1) F. Kudo, T. Akitomi, and N. Moriwaki, “An Artificial

Intelligence Computer System for Analysis of Social-

Infrastructure Data,” IEEE conf. Business Informatics (CBI)

(Jun. 2015)

(2) J. Kimura et al., “Framework for Collaborative Creation with

Customers to Improve Warehouse Logistics,” Hitachi Review

65, pp. 873–877 (Mar. 2016).

(3) Hitachi News Release, “Development of Artificial Intelligence

issuing work orders based on understanding of on-site kaizen

activity and demand fluctuation,” (Sep. 2015), http://www.

hitachi.com/New/cnews/month/2015/09/150904.html

with an understanding of the worksite improvements

and environmental changes. A demonstration test

of picking work improvement was conducted in a

distribution warehouse, and a work reduction of 8%

was obtained as the result. In the future, Hitachi will

work on further generalization of this technology, and

on expanding its application into other fields such as

manufacturing and distribution.

Tomoaki Akitomi

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi Ltd. He is currently engaged in the research and development of human behavioral science and AI. Mr. Akitomi is a member of the JSAI.

Atsushi Miyamoto, Ph.D.

Center for Exploratory Research, Research & Development Group, Hitachi Ltd. He is currently engaged in the research and development of big data analytics and AI.

Ryuji Mine

Center for Exploratory Research, Research & Development Group, Hitachi Ltd. He is currently engaged in the research and development of AI, evolution science, and learning science. Mr. Mine is a member of the Institute of Electronics, Information and Communication Engineers (IEICE), JSAI, and the IPSJ.

Junichi Hirayama

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi Ltd. He is currently engaged in the research and development of big data analytics, AI, and technology for improving logistics operations. Mr. Hirayama is a member of The Japanese Society for Artificial Intelligence (JSAI) and the Information Processing Society of Japan (IPSJ).

Fumiya Kudo

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi Ltd. He is currently engaged in the research and development of AI and big data analysis based on statistical approaches. Mr. Kudo is a member of the JSAI.

ABOUT THE AUTHORS


- 61 -

Featured Articles

Utilization of AI in the Water Sector

Case Study of Converting Operating History Data to Values

Ichiro Embutsu, Dr. Eng.

Koji Kageyama, Dr. Eng.

Satomi Tsuji

Norihiko Moriwaki, Ph. D.

Yukiko Ichige

OVERVIEW: Global demand for water has been increasing as urban economic activity expands. A significant amount of energy is required to obtain clear water, and the reduction of this energy requirement is a major concern for water utilities. Since the qualities of raw water vary greatly in water processing systems, such systems are constructed out of combinations of multiple unit processes designed to handle each of these qualities. For this reason, there are many cases where explicit models based on physical and chemical phenomena are not sufficient for implementing appropriate operation and control, and so expectations have grown for the utilization of implicit knowledge that is inherent in the operating history data. Hitachi is deploying its Hitachi AI Technology/H technology to social infrastructure, and will continue contributing to water supplies that are both safe and reliable by proactively applying the system to water treatment systems.

INTRODUCTION

THE global water environment market shows promise and is poised to grow from 36.2 trillion yen in 2007 to 86.5 trillion yen in 2025(1). Of this, just under 90% is comprised of water supply and sewage systems, between 40% and 50% of which is management and operation. The over 10% remaining is taken up by seawater desalination, industry, and recycling segments which are expected to grow tremendously.

Hitachi is promoting its Social Innovation Business in a bid to reform social infrastructure by fully utilizing information and communication technologies (ICT), and its provision of solutions for the water industry is playing a key role in these efforts. Specifically, by supplying products, systems, and services, Hitachi is working to provide solutions for customers’ issues in areas such as water conservation, flood control, water supply and sewage systems, securing water resources (desalination and water recycling), and wastewater treatment, etc.

The biggest concern of water utilities introducing solutions such as these is how to minimize business costs while maintaining the regulated level of water quality. Most of all, along with streamlining and the reduction of manpower requirements related to the operation of water processing systems, Hitachi

is giving priority to energy-saving solutions as a major focus of its research and development, since expectations for them are high.

The technology of artificial intelligence (AI) has seen new progress in recent years, and is expected to be effectively applicable to water treatment systems as well. This article introduces these efforts.

ATTEMPTS AT UTILIZING AI IN THE WATER SECTOR

Water supply and sewer systems, seawater desalination systems, and other water treatment systems differ greatly from other typical industrial systems in that fluctuations cannot be avoided in the qualities of the raw water that they process (river water, sewage, seawater, etc.). In order to deal with these changing raw water qualities, water treatment systems are comprised of multiple unit processes including sedimentation, biological treatment, membrane filtration, and others. The physicochemical phenomena that occur within these unit processes are systematized and formulated based on previous knowledge, and although this is controlled automatically to a great extent, it depends on the know-how and skill level of operators. Sometimes flexible responses are required, and it is important to provide solutions that can deal with these types of cases.

140 Case Study of Converting Operating History Data to Values

- 62 -

The history of attempts to apply AI is relatively long in the water sector, and even Hitachi experimented with applying the know-how of operators involved with water treatment systems and the causal relationships inherent in operating history data to operation and control in the 1990s. For instance, Hitachi applied fuzzy logic-based expert systems, which can be considered to be one type of AI in the broader sense, along with neural networks to coagulant chemical injection operations at water treatment plants, and demonstrated at the actual plant level that these technologies can be used to handle operation during both normal and abnormal situations, including high-turbidity raw water(2), (3).

Due to advancements in AI technology and improvements in machine power that have been occurring in recent years, the environment is in place for utilizing more massive amounts of operating history data than before. The next chapter describes in detail how Hitachi is considering the application of AI technology with a focus on the seawater desalination sector.

CASE STUDY EXAMINING DESALINATION SYSTEM OPERATION AND CONTROL

Target System: Water Desalination & Reclamation SystemFig. 1 shows an example of a representative process flow using the “water desalination & reclamation system” integrated seawater desalination and sewage

treatment system. This system recycles treated sewage and other types of water while at the same time utilizing the concentrated water (brine) generated in the final filtration process as dilution water for its seawater desalination system, thereby achieving low-cost desalination that saves energy while reducing the burden placed on the environment. The system can be broadly divided into a sewage and industrial drainage recycling system and a seawater desalination system. The sewage and industrial drainage recycling system biologically treats the sewage in a membrane bioreactor (MBR), and then filters the output through a sewer system reverse osmosis (RO) membrane device to get water for reuse. With this water, it is possible to achieve water quality that is at the level of drinking water or industrial water.

The seawater desalination system filters seawater through ultrafiltration (UF) and then mixes the resulting water with the concentrated water in the sewer system RO membrane device, and then filters the output through a seawater RO membrane device to get purer water, after which the seawater RO membrane device’s concentrated water is released into the ocean as drainage water. When compared with general seawater desalination systems, this system offers the following four advantages:(1) It can effectively utilize the sewer system RO membrane device’s concentrated water output from the sewage recycling system, thereby reducing the amount of drainage water.

Fig. 1—Water Desalination & Reclamation System Process Flow Example.The figure shows a representative configuration example integrating seawater desalination and the recycling of sewage and other input.

MBR

Sewage and industrial drainage recycling system

Sewage

Seawater

Industrial drainage

Seawater desalination system

UF

UF

Sewer system RO

Seawater system RO

Discharged water

Water for reuse(for industrial use)

Salt, ions, and other impurities are removed

Salt, ions, and other impurities are removed

Low-pressure pumpOrganic system

contaminants are removed

Bacteria and other particles removed

Water intake amount reduced due to use of low-concentration concentrated water

Cost of powering pump is reduced by diluting the seawater

Salinity is reduced to the level of seawater

Medium-pressure pump

Low-concentration concentrated water

MBR: membrane bioreactor RO: reverse osmosis UF: ultra filtration


- 63 -

(2) It can reduce the quantity of seawater intake

required to produce a given amount of production

water (fresh water), thereby reducing the size of water

intake facilities while at the same time lowering the

cost of power required for water intake.

(3) By mixing seawater with the concentrated water

from the sewer system’s RO membrane device, it

reduces the osmotic pressure of the treated water in

the seawater system’s RO membrane device, thereby

lowering the cost of powering the filtration pumps

necessary for desalination.

(4) The salinity of the concentrated water of the

seawater system’s RO membrane device is reduced

to roughly the level of seawater.

On the other hand, a seawater desalination system

based on RO membrane devices does suffer from the

same problem of fouling, which is widely known to

increase the cost of power to run the filter and to

decrease equipment utilization. Although there is

established knowledge regarding the mechanisms

that cause fouling(4), at present, effective suppression

methods still rely in part on trial and error. It is

for this reason that Hitachi attempted to acquire

knowledge regarding operation and control methods

that can suppress fouling by applying AI technology

to operating history data from the past.

Applied AI Technology: Hitachi AI Technology/HIn order to acquire new knowledge regarding the

suppression of fouling, Hitachi is applying AI

technology it has developed called Hitachi AI

Technology/H (hereafter referred to as H). This

technology offers functions that exhaustively derive

and visually represent correlations from large amounts

of numerical data (indices) generated by combining

huge amounts of data(5). In this way, the system can

extract from among the many different types of indices

those that have useful correlations with objective

variables, and use this information to create specific

measures that are highly effective with respect to the

objective variables.

By inputting and analyzing operating history

data from the water desalination & reclamation

system, Hitachi is attempting to come up with

control methods based on new causal relationships

that have been overlooked in the past. This includes,

for example, the expected ability to extract candidate

process parameters that correlate meaningfully with

the seawater system RO membrane device’s inlet

pressure, which increases when fouling occurs.

Used Data and Analytical MethodsThis analysis used operating history data acquired

through “Water Plaza Kitakyushu,” which was

contracted to the Global Water Recycling and Reuse

Solution Technology Research Association (GWSTA)

as a project for the New Energy and Industrial

Technology Development Organization (NEDO).

Hitachi used the acquired data by selecting data at one-

hour intervals without irregular operations in order to

create an analytical data set comprised of one objective

variable and 43 explanatory variables.

The analytical method was implemented based on

the four-step process shown in Fig. 2, with the goal

of extracting the influencing factors affecting inlet

pressure at the seawater system RO membranes that

could indicate the degree of fouling, and which could

be used as a basis to consider methods of fouling

suppression control. For this reason, the analysis was

conducted using the seawater system RO membrane

inlet pressure as the objective variable.

First, Hitachi extracted influencing factors from

this data set using H, and then extracted explanatory

variables (influencing factors) based on these results

that had meaningful correlations with the objective

variable. The relationships between the extracted

influential factors and the objective variable were then

visualized using a network format. Phenomenological

pre-existing knowledge regarding fouling was then

used to extract control indices and devise control logic.

Finally, the devised control logic was used to estimate

the expected fouling suppression benefit, and both

feasibility and practicality were evaluated.

Analysis Results and Consideration of Control MethodsThe results of the analysis using H with seawater

system RO membrane inlet pressure selected as the

Step 1: Extract influential factors with respect to the target of control using Hitachi AI Technology/H

Step 2: Visualize correlations between indices in a network format

Step 3: Extract control indices using pre-existing knowledge and propose control logic

Step 4: Estimate the benefit of proposed control logic (evaluate the feasibility and practicality)

Fig. 2—Analysis Flow for Consideration of Control Logic.H was applied in a four-step analysis.


- 64 -

objective variable produced correlation coefficients with 43 explanatory variables. Of these variables, Hitachi focused on one related to the quality of the sewer system RO membrane’s concentrated water, which accounted for approximately half of the water supplied to the seawater system’s RO membranes. As shown in the scatter plot in Fig. 3, a direct correlation existed with respect to the objective variable. Increasing electrical conductivity is thought to be associated with increases in salinity and other factors, and the connection with the increasing inlet pressure of the seawater system’s RO membranes agrees with the phenomenological analysis as well.

Furthermore, Fig. 4 shows the results of visualizing the relationships of other variables with the conductivity of the concentrated water from the sewer system’s RO membranes, based on the network analysis function applied between indices. When viewed from the perspective of changing conductivity and then verifying results, the flow rate of water through the sewer system process was derived as a variable factor. The related operating history data also showed that specific increases in conductivity were triggered by changes in the flow rate of water through the process.

Of the knowledge attained from the analysis described above, the knowledge of what was seen as an effective method for suppressing increases in the inlet pressure of the seawater system’s RO membranes, which was the objective variable (causal relationships between variables), can be summed up with the following inferred causal relationship: suppressing the electrical conductivity of the concentrated

water from the sewer system’s RO membranes will suppress the electrical conductivity of the mixture of the concentrated water from the sewer system’s RO membranes with seawater, which will then suppress the osmotic pressure of the mixed water, thereby suppressing the inlet pressure (absolute value) of the seawater system’s RO membranes and/or the inlet pressure’s increase over time. A conceivable control plan based on this is, (1) controlling the mixed water electrical conductivity. Conceivable subordinate control plans include, (2) controlling the electrical conductivity of the sewer system RO membrane’s concentrated water, (3) controlling the blend ratio, and (4) controlling the seawater intake. The specific control methods derived are shown in Fig. 5.

Hitachi estimated the suppression benefit with respect to increases in inlet pressure at the seawater system’s RO membranes based on these control methods (details such as the method of estimation are omitted here). Estimates showed that operation and control that restrained the flow rate of the sewer system process for the approximately ten days of the evaluation period resulted in a total pressure increase (integrated value) of approximately 6% of the 1.47 MPa that would occur without this control, producing a figure of only 0.09 MPa. Analysis of a breakdown of this benefit showed a direct benefit

3,7003,6753,6503,6253,6003,5753,5503,5253,5003,4753,4503,4253,4003,3753,3503,325

1,750 2,000 2,250 2,500 2,750 3,000

Design index: electrical conductivity of sewer system RO concentrated water

Since conductivity is proportional to salinity, the higher it is, the higher the RO membrane’s inlet pressure becomes (this trend matches the phenomenological conclusion as well).

Obj

ectiv

e in

dex:

sea

wat

er s

yste

m R

O m

embr

ane

inle

t pre

ssur

e

3,250 3,500 3,750 4,000 4,250

Fig. 3—Correlation Chart between Water Quality Parameters Extracted by H and Objective Variables.The results imply a trend where the higher conductivity is, the higher RO membrane inlet pressure becomes. Fig. 4—Example of Visualization in Network Format of

Correlation between Indices Related to Sewer System RO Membrane Concentration Water Electrical Conductivity Based on Results of H Analysis.The controllable factors can be extracted by visualizing correlative relationships.


- 65 -

In addition to seawater desalination plants, this

technology can be applied to any other plant as long

as past operating history data is available. Hitachi will

continue working to expand the application of this

technology to other plants, including water supply and

sewer systems as well as other water treatment plants.

The authors would like to thank GWSTA for

its permission and cooperation in providing the

operating history data used for the evaluation tasks

described above.

REFERENCES(1) Ministry of Economy, Trade and Industry, “Challenges and

Specific Measures for International Development of the Water

Business,” (Apr. 2010) in Japanese.

(2) I. Embutsu et al., “Rule Extraction from Neural Network—

Application for the Operation Support of Coagulant Injection

in Water Purification Plant—,” Transactions of the Institute

of Electrical Engineers of Japan D, Vol.111, No.1, pp. 20–28

(Jan. 1991) in Japanese.

(3) I. Embutsu et al., “Integration of Multi AI Paradigms for

Intelligent Operation Support Systems,” Water Science &

Technology, Vol.28, No.11 (Dec. 1993)

(4) J. S. Vrouwenvelder et al. “Biofouling of Spiral Wound

Membrane Systems,” Journal of Membrane Science, Vol.346,

Issue.1 (Jan. 2010)

(5) N. Moriwaki et al., “AI Technology—Achieving General-

Purpose AI that Can Learn and Make Decisions for Itself,”

Hitachi Review 65, pp. 35–39 (Jul. 2016).

of approximately 3% due to a reduction in osmotic

pressure through lowered conductivity. The benefit of

suppressing the increase in irreversibility of filtration

pressure (that is, fouling suppression) was also shown

to be approximately 3%.

Since the cost of powering a high-pressure pump

can be considered to be proportional to the filtration

pressure, this means that the cost of power is expected

to be reduced by approximately 6%. At present, when it

comes to the operation of seawater desalination plants,

a variety of different on-site efforts have incrementally

reduced operating costs a steady rate of several tenths

of a percent. The proposed control methods are judged

as providing a significant benefit in terms of reducing

running expenses without generating additional costs

or requiring new equipment or chemicals.

CONCLUSIONS

The water desalination & reclamation system is a

process that integrates both sewage recycling and

seawater desalination, and Hitachi is aiming to expand

its adoption both domestically and internationally.

By applying Hitachi AI Technology/H to the fouling

problem shared by all water treatment systems that

use filtration membranes, it was possible to extract the

knowledge necessary for considering control methods.

N

N

Y

Y

Repeat over predetermined range

Repeat over predetermined range

AElectrical conductivity of sewer

system RO concentrated water > predetermined value 1

BMixed water electrical conductivity >

predetermined value 2

A-1 : Sewage treatment flow rate

*If control is not possible in A-1

*If control is not possible in B-1

B-1 : Seawater blend ratio

B-2 : Halt seawater intakewhile seawater electrical conductivity >

predetermined value 3

Fig. 5—Control Method Flow for Suppressing Fouling.This control method was derived based on knowledge extracted using H.


- 66 -

Koji Kageyama, Dr. Eng.

Process Engineering Research Department, Center for Technology Innovation – Materials, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of water treatment and control systems for water supply and sewage. Dr. Kageyama is a member of EICA.

Norihiko Moriwaki, Ph. D.

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of human information systems and artificial intelligence. Dr. Moriwaki is a member of The Institute of Electronics, Information and Communication Engineers (IEICE), The Japan Society for Management Information (JASMIN), and the Association for Information Systems (AIS).

Yukiko Ichige

Business Development Department, Global Water Solutions Division, Water Business Unit, Hitachi, Ltd. She is currently engaged in business development and task management of water solutions for the global market. Ms. Ichige is a member of The Robotics Society of Japan (RSJ).

Ichiro Embutsu, Dr. Eng.

Process Engineering Research Department, Center for Technology Innovation – Materials, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of water treatment and control systems for water supply and sewage. Dr. Embutsu is a member of The Society of Environmental Instrumentation Control and Automation (EICA), the Japan Society on Water Environment (JSWE), and The Institute of Electrical Engineers of Japan (IEEJ).

Satomi Tsuji

Global Center for Social Innovation – Tokyo, Research & Development Group, Hitachi Ltd. She is currently engaged in research into the use of human big data for organizational management. Ms. Tsuji is a member of the Society of Project Management (SPM).

ABOUT THE AUTHORS


- 67 -

Featured Articles

Utilization of AI in the Manufacturing Sector

Case Studies and Outlook for Linked Factories

Naohiko Irie, Dr. Eng.

Hiroto Nagayoshi

Hikaru Koyama

OVERVIEW: Utilization of the IoT is steadily advancing in the manufacturing sector. In response to this trend, Hitachi is working to provide new industry solutions based on the symbiotic autonomous decentralization concept for achieving overall optimization of activities and for creating new business by linking various systems and stakeholders together. To make new industry solutions a reality, it will be necessary to have on-site sensing, to provide an infrastructure for collecting and archiving big data, to analyze and plan countermeasures, and to give feedback to sites. This article describes advanced on-site sensing, integrated analysis of a variety of on-site data for analyzing and planning countermeasures, and the utilization of AI technology for them.

INTRODUCTION

IN recent years, utilization of the Internet of Things (IoT) has resulted in increasingly active trend towards a new evolution of the manufacturing industry. The Industrial Internet Consortium (IIC), which includes General Electric Company (GE) as one of its founders, and the government-led Industrie 4.0 were established in the USA and Germany, respectively. Both of these organizations are engaged in the formation and standardization of a new ecosystem that involves the manufacturing industry and information technology (IT) industry.

Hitachi has a track record for building large-scale control systems in diverse fields, such as energy, transportation, and water supply and sewerage, in addition to production management systems and control systems designed for various manufacturing industries including steelmaking, automobiles, and medicine. Hitachi is leveraging its knowledge of systems such as these to link diverse systems. It is advocating the “symbiotic autonomous decentralization” concept(1) to provide value gained from this to the manufacturing industry and the social infrastructure sector, and to promote new growth.

Symbiotic autonomous decentralization allows sensing of a site’s various statuses (Sensing), analysis of issues and planning of countermeasures based on various collected and archived information (Thinking),

and feed back of the results obtained to the site (Acting), thus enabling the optimization of value chains inside and outside of the factory.

This article describes new industry solutions that will be achieved by such symbiotic autonomous decentralization, and the machine learning technology and artificial intelligence (AI) that form the core of these systems.

LINKED FACTORIES AND NEW INDUSTRY SOLUTIONS ACHIEVED BY SYMBIOTIC AUTONOMOUS DECENTRALIZATION

Conventional optimization at production sites is limited to analysis at the individual system level and improvement of each site based on this, and the effectiveness of such improvements is becoming saturated. This is why symbiotic autonomous decentralization is being applied to achieve linked factories with a view toward optimizing activities across multiple systems and the creation of new value chains. To make this a reality, Hitachi is collecting and archiving information from other related systems in addition to the information gained from manufacturing systems, and is also further analyzing the information of other factories that are being expanded globally, planning countermeasures, and giving feedback to sites with the aim of optimizing activities overall (see Fig. 1).

146 Case Studies and Outlook for Linked Factories

- 68 -

Depending on the information to be combined, the following types of solutions are possible.(1) Improved energy productivity

It is predicted that energy costs will fluctuate considerably as a result of the liberalization of electric power, violent fluctuation of the crude oil price, utilization of renewable energy, and other factors, and improvements in productivity to correspond with this fluctuation will be required. For this reason, Hitachi is collecting data on the status of energy consumption in addition to detailed production site information from sources such as manufacturing execution systems (MESs) and points of production (POPs), is conducting analysis regarding the relation between energy and productivity, and is formulating energy procurement plans and production plans that will lead to the reduction of overall energy costs and peak shaving. Formulated plans are fed back to an MES and the optimum production plan is executed so that renewable energy in the factory is used effectively and energy costs are reduced.

(2) Supply Chain Management (SCM) coordinationTo achieve optimum operation of factories that

are expanding globally, site information from MESs and inventory control systems is analyzed, and high-precision management indicators are evaluated by a business value simulation tool that factors in the logistical status of each region. These indicators are then used to determine the optimum production plan for each site, distribution routes, stock quantities, etc., and the production plan is then executed based on these parameters. Also, since on-site information is analyzed in real time, production capacity and inventory is adjusted to accommodate unforeseen circumstances.(3) Global quality management/improvement

Ensuring consistent quality at a high level between factories that are expanding globally not only results in cost reductions but also reduces the risk of product recalls. To accommodate this, worker actions are sensed by video analysis as well as detailed manufacturing information from MES and POP. The information that is gained is archived and analyzed so that factors that

Managers, business planners

Analysis

Data collection and archiving

Symbiotic autonomous decentralized platform

Acting

Insight

Sensing

Warning signdetection

Cause analysis

Plan countermeasures

SchedulingOT×IT

Security

Operationalsuggestions

ThinkingThinking

Thinking

Feedback to site


Acting

Insight

Sensing


Analysis



Acting

Insight

Sensing


Cause analysis


SchedulingOT×IT

Security


ThinkingThinking

Thinking

Analysis

Global bases Other industries

Energy systems Security systemsLogistics systems

Manufacturing site systems Quality control systems



ActingSensing


Cause analysis


Feedback to site

Feedback to site

SchedulingOT×IT

Security


Thinking

ActingSensing

Thinking

ActingSensing

Thinking

Thinking

Site machinery and equipment On-site systems Information systems

Fig. 1—Symbiotic Autonomous Decentralization Concept.Activities can be optimized in terms of management considerations and new value chains created by collecting and archiving data from a variety of on-site systems, analyzing it and planning countermeasures, and providing information back to the site as feedback.

OT: operational technology IT: information technology


- 69 -

degrade quality can be extracted and countermeasures

can be planned. Hitachi will introduce new analytical

methods such as AI since the factors that affect quality

are complex and are predicted to be wide-ranging.

Obtained countermeasures will be fed back to workers

using augmented reality (AR) as well as MESs and

other control systems.

(4) Business Continuity Plan (BCP) support

In recent years, cyber-attacks have increased

the risk of factory operations being shut down. As

a result, the impact on business when an incident

occurs must be minimized. Monitoring information,

detected illegal access to control systems, and detected

viruses, as well as detailed information collected from

manufacturing systems, is collected from each base.

When an incident such as a virus infection is reported

by the security monitoring center, the impact of the

virus itself, the scope of impact from the viewpoint

of control system configuration, and the impact on

business operations from the viewpoint of production

status are judged as a whole, and plans, for example,

for setting the system offline, eradicating the virus, or

adjusting production capacity are formulated.

UTILIZING AI TO ACHIEVE NEW INDUSTRY SOLUTIONS

Technology for (1) conducting advanced sensing of

site information and (2) analyzing a variety of data to

plan countermeasures will be the keys to achieving the

solutions described above. With regard to advanced

sensing, video analysis technology is gradually being

put to use in obtaining birds-eye-view like information

such as people’s movements, and learning functions,

AI, or the like for comprehending the meaning of

captured video are anticipated.

For the analysis and planning of countermeasures,

it is strongly desirable to utilize AI from the viewpoint

of analyzing diverse data in an inter-disciplinary

manner to obtain new knowledge.

The following describes actual case studies where

AI was utilized.

Application of Machine Learning to the Recognition of Work ActivitiesHitachi is developing a sensing technology for

recognizing the movements of factory workers and

for detecting worker movements that deviate from a

predetermined standard range (i.e. abnormal operation)

for the purpose of improving product quality. The

following describes the machine learning that forms

the core of this sensing technology.

First, a motion camera is used to recognize worker

movements. This allows joint position information

(e.g. wrists, elbows, shoulders) to be obtained from

a worker’s 3D shape. Machine learning is then used

based on the obtained joint position information to

recognize the worker’s movements.

Fig. 2 shows an overview of the abnormal operation

detection algorithm. In the preprocessor, noise in

the joint information is removed by smoothing, and

information that is not directly related to work, such

as arm length or leg length, is canceled out through

normalization. The feature value extractor extracts

feature values, which are pieces of information that

represent movements. In the classifier, combinations

of feature values are selected according to the kind of

work in which abnormal operation is to be detected.

Preprocessor

Smoothing

Head movement extraction

Waist movement extraction

Feature value

selection

Judgment of abnormal operation in two-handed work

Judgment of abnormal operation in single-handed work

Judgment of abnormal operation in whole-body work

Standard operation model

Hand movement extraction

Normalization

Joint position

information

Judgment result• Abnormal

operation type• Degree of

abnormal operation

Feature value extractor Classifier

Fig. 2—Overview of the Abnormal Operation Detection Algorithm.Machine learning is applied in each of the abnormal operation judgment processes in the classifier.


- 70 -

And, the presence of each abnormal operation is

judged (abnormal operation judgment) by making a

statistical comparison with a standard operation model.

Machine learning can be put to effective use by

the feature value extractor and classifier. In particular,

when there is a large amount of data, both can be

optimized using a technology called “deep learning,”

in other words, automatic design is now becoming

possible. However, when a large amount of data

cannot be obtained, prior human knowledge is used to

execute optimization up to the feature value extractor

and to design the combinations of feature values.

With the proposed techniques, work movements were

decomposed down to their motion elements based on

the observation and investigation of on-site work, and

those elements were taken as prior knowledge in the

design of the feature value extractor and combinations.

Up to this point, the process is qualitative design,

which can be understood by humans.

On the other hand, with abnormal operation

classification, the question is: what kind of judgments

are to be made about the feature values, namely,

quantitative design is required. Machine learning is

generally excellent for quantitative design.

There are two types of machine learning, supervised

and unsupervised. With supervised machine learning,

two types of training samples are used, normal

operation samples and abnormal operation samples.

When a training sample is input, the abnormal operation

classifier is optimized so that the correct judgment

is made. However, collecting abnormal operation

samples is difficult, for two reasons: there are few

abnormal operations and there are countless variations.

For this reason, unsupervised training, which uses

only normal operation samples to perform learning, is

used. As a result, the standard operation model can be

estimated as a probability distribution.

Fig. 3 shows a conceptual diagram of a probability

distribution. Judgment results are expressed by single

points within this distribution, and the further the

judgment is from the center of the distribution, the

stronger the degree of abnormal operation is going

to be. For example, the center is the operation that is

taken as the model, and the peripheral area around the

center is a normal operation. If we move further away

from the center, operations that require caution or

abnormal operations can be demarcated as caution or

abnormal, respectively. It has been demonstrated that

operations that different from the norm can be actually

extracted by using the standard operation model that

has been learned.

Utilizing AI for Analyzing Site DataGenerally, distributed control systems (DCSs), which

are in charge of the direct control of facilities, and

MESs, which are in charge of production and quality

management, are installed at productions sites. These

systems collect enormous volumes of data relating

to manufacturing facilities, production processes,

and product quality every day. This data is mainly

archived as numerical values and have been analyzed

and utilized by statistical quality control techniques

up to now. However, the rapid increase in the volume

of archived data has made it humanly difficult to

search for relationships between that data and key

performance indicators (KPIs), such as quality and

non-defective product ratio at production sites, and

identify causes.

For this reason, Hitachi has been implementing

analysis using Hitachi AI Technology/H (hereafter

referred to as H), its own proprietary artificial

intelligence technology(2). When a KPI and data

potentially related to the KPI (explanatory parameters)

are input, H automatically generates feature values

from the explanatory parameters, comprehensively

calculates the correlation with KPI, and outputs

statistically significant feature values (see Fig. 4). The

characteristic of H here is that it generates combinations

of explanatory parameters as feature values.

For example, with product manufacturing in a

discrete system, assume manufacturing equipment

X and Y, for which the processing values fall roughly

within the ranges 1.0 to 4.0 and 5.0 to 10.0, respectively.

Dimension 2 offeature value

Dimension 1 offeature value

Abnormal

Model

Normal

Caution

Fig. 3—Conceptual Diagram of Probability Distribution Showing a Standard Operation Model.A standard operation model is estimated as a probability distribution. Judgment results are expressed by single points within this distribution, and the further the judgment is from the center of the distribution, the stronger the degree of deviation will be.


- 71 -

With production in a continuous system, each piece

of equipment and each process generally has designed

control values, and as long as these control values fall

within the control ranges at each step of production

there is no problem. However, production processes

fluctuate on a daily basis, and appropriate operating

conditions for the production processes keep changing

because of the wearing and degradation of parts, and

before and after facility maintenance. Furthermore,

operating conditions and control conditions must be

reviewed when production is switched over to new

products. H is effective for discovering the appropriate

operating conditions for responding to changes such

as these and for gaining new awareness based on large

amounts of data. For example, in processes that are

comprised of a total of ten steps, combinations of data

across processing in the first and third steps can be

analyzed. The ability to analyze a wide breadth of data

across steps and processes in this way is an advantage

of H, and this can be used as an opportunity to make

workers and managers at production sites, who tend to

focus on the management of facilities and processes,

more aware.

CONCLUSIONS

This article described linked factories and new

industry solutions based on the symbiotic autonomous

decentralization concept, as well as the utilization of

AI that will be key in achieving this.

Although the article dealt mainly with utilization

of AI geared towards improving quality, in the future,

Hitachi intends to increase the number of case studies

where AI is utilized in other solutions, and to apply AI

technology to production and engineering sites while

advancing the verification of its effectiveness with a

view to supporting manufacturing innovation.

REFERENCES(1) N. Irie et al., “Information and Control Systems—Open

Innovation Achieved through Symbiotic Autonomous

Decentralization—,” Hitachi Review 65, pp. 13–19 (Jun.

2016).

(2) K. Yano, “Invisible Hand of Data: The Rule for People,



When the data from manufacturing equipment X and Y

and the KPI (here: production volume) are input to H,

an enormous volume of feature values is generated, the

correlation with KPI is comprehensively analyzed, and

statistically significant feature values are extracted.

Fig. 5 shows an example of the analysis. The

product group that satisfies the feature value (X:

1.0 to 1.5 and Y: 9.5 to 10.0) discovered by H is

shown as Applicable, and the average value of that

KPI (production volume) was 100 (relative value).

Whereas, the product group that does not satisfy

the feature value (where either X: 1.0 to 1.5 or Y:

9.5 to 10.0 deviate outside this range) is shown as

“Not applicable,” and the average value of that KPI

(production volume) is 94. It can be seen that there

is a difference of 6% in production volume between

when product manufacturing satisfies the feature value

and when it does not. In other words, an improvement

in production volume of 6% can be anticipated by

controlling manufacturing equipment X and Y so that

the feature value is satisfied.

KPI

Not applicable Applicable

6%

85

90

95

100

105

110

Fig. 5—Example of Results Discovered by H (Relative Values taking Applicable to be 100).“Applicable” indicates a group that satisfies the discovered feature values, and “Not applicable” shows a group that does not.

(1) KPI(target indicator)

(3) Feature value

Automatic extractionof correlation


(2) Production site data

Fig. 4—Diagram Illustrating Input/Output of H.When (1) KPI and (2) KPI-related production site data are input, then (3) multiple correlated feature values are output.



- 72 -

Hiroto Nagayoshi

Media Systems Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of computer vision and pattern recognition. Mr. Nagayoshi is a member of the IPSJ, the Institute of Electronics, Information and Communication Engineers (IEICE), and the IEEE.

Naohiko Irie, Dr. Eng.

Center for Technology Innovation – Controls, Research & Development Group, Hitachi, Ltd. He is currently engaged in the research and development of control systems and platforms. Dr. Irie is a member of the Information Processing Society of Japan (IPSJ).

Hikaru Koyama

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi, Ltd. He is currently engaged in research into applying AI for the manufacturing industry. Mr. Koyama is a member of The Japan Society of Applied Physics (JSAP), The Surface Science Society of Japan (SSSJ).

ABOUT THE AUTHORS


- 73 -

Featured Articles

Advanced Research into AI

Debating Artifi cial Intelligence

Kohsuke Yanai, Ph.D.

Yoshiyuki Kobayashi, Ph.D.

Misa Sato

Toshihiko Yanase, Ph.D.

Toshinori Miyoshi, Ph.D.

Yoshiki Niwa, Ph.D.

Hisashi Ikeda, Ph.D.

OVERVIEW: This article describes a debating AI that can debate decision-making matters with humans. Given a discussion issue, the system outputs short argument scripts based on multiple viewpoints. The argument scripts are summarized information with a length of about 10 sentences, which is composed of evidence, analytics, and opinions obtained from large amounts of document data. They can be thought of as a consolidated essence for decision-making. The debating AI can automatically investigate business opportunities and risks on a regular basis from multiple points of view. With the aim of creating an enterprise IT system that drives innovation in organizations, Hitachi is accelerating the development of the debating AI through open collaborative creation.

INTRODUCTION

HITACHI has been working on developing a debating artificial intelligence (AI) that can debate decision-making matters with humans. Given a discussion issue and a stance, the system provides useful information for making decisions, such as grounds and counterexamples. For the issue “Should we use electric vehicles as office cars?” for example, it answers “Electric vehicles are environmentally conscious” when asked for a positive stance, or “Electric vehicles are expensive” when asked for a negative stance. Next, it provides evidential information extracted from news articles, white papers, research reports, and other material.

Fig. 1 shows an overview of the debating AI. On the present computing platform, it composes three opinions within 80 seconds. Each generated opinion includes a different viewpoint and is based on the analysis of about 10 million English news articles. The purpose of using the AI system is to explore discussion issues by “debating” with humans, and to induce evidence-based decisions.

The rest of this section explains the background of the development. Hitachi assumes that data analytics technologies will evolve through the three phases illustrated in Fig. 2. While the second phase, optimization using big data, has been well-studied, the focus here is on the third phase. In today’s fast-

moving world, companies need to continuously produce innovative services and value. R. G. McGrath, a professor at Columbia Business School, argues that competitive advantage is becoming harder to maintain. In an era of transient competitive advantage, companies need to build and manage multiple innovations to keep exploiting transient advantages(1). She also makes the case that the process of developing innovations should not be experimental, but central to corporate strategy. In such an environment, strategic planning should take on a greater importance within enterprises. Therefore, there is value in being able automatically to find business opportunities and risks on a regular basis from multiple points of view through integrating various information resources, such as news articles, company databases, user reviews, white papers, and research reports.

Hitachi has been developing the debating AI with the aim of creating an enterprise IT system that automatically investigates business opportunities and risks. Believing it important to provide counterexamples that represent the risks associated with decisions, the focus is on techniques for composing counterarguments.

The following section describes the value provided by the debating AI. Subsequent sections describe an evidence recognition technique and social implementation through open collaborative creation.

152 Debating Artifi cial Intelligence

- 74 -

VALUE PROVIDED BY DEBATING AI

As shown in Fig. 1, the debating AI generates short scripts that describe opinions for a given discussion issue. The generated short scripts are called “arguments.” The definition is as follows.

Argument: A summarized text, consisting of about 10 sentences, which is composed of evidence, analytics, and opinions from large amounts of document data.

The authors estimate that reading a generated argument is roughly equivalent to reading 400 sentences

obtained by querying the discussion topic using a traditional text search technique. The intention is that the arguments will provide a consolidated summary of the material for decision-making.

The arguments are likely to be of most use when people need to read large amounts of document data to investigate a decision. Typical cases are as follows.(1) Analysis of facilities and policies

Consider the case of investigating past examples prior to building a new casino. It would be desirable to gather information from various accessible resources

1. Agility(Monitoring)

2. Efficiency(Prediction/

Optimization)3. Risk aversion

(Evidencing)

3. Opportunity(Exploring)

Managementinformation

Business intelligence

Big data

Machine learning

Knowledgeinformation

Text understanding

Fig. 2—Values of Data Analytics.While past studies have dealt with the second phase of optimization using big data, it will also be important to automatically investigate business opportunities and risks on a regular basis.

Given a discussion issue, the system composes opinions by integrating news articles, white papers, and

research reports.

News articles

White papers Research reports Provide opinions from multiple viewpoints

Discussion issue

Combine Evidence and counterexamples

Fig. 1—Overview of Debating AI.The screenshot is from the interface of the debating AI. When a discussion issue “We should ban casinos” is given, it composes opinions by integrating various information resources. The above example illustrates that the system provides opinions based on three viewpoints: addiction, crime, and economy.

AI: artificial intelligence


- 75 -

that describe similar cases from the past about matters such as how much employment was actually created, how the crime rate changed, whether the economy improved or worsened, and the response of residents in the years following the opening of a new casino. Such information would be available in white papers and news articles. In general, investigations of this nature are required whenever companies introduce new facilities or policies.(2) Analysis of companies or organizations

It can be useful for investigating information about customers or business partners, such as what kind of products a company has launched in the past, the public response to the product, what kinds of problems the company is facing, and whether the company is engaged in any leading-edge programs with foreign governments. In general, numerical data such as profit margins, growth rates, and stock prices are the most important. However, when people have to make important decisions, they also tend to refer to textual information of the sort described above. In some cases, a decision may be changed based on negative evidence despite numerical data supporting the decision. Requirements of this nature tend to occur when making investment decisions or selecting a business partner.(3) Analysis of locations or markets

When a company plans to expand its business in a new location, it needs to find out about any local security risks in the area, economic trends, the state of social infrastructure, and information about social problems. It also needs up-to-date information about residential interests and important local government plans. While numerical data is available on things like population changes and industrial structure, this can be complemented by textual information.

CORE TECHNICAL COMPONENTS OF DEBATING AI

The debating AI consists of the following technical components(2), (3).(1) Understanding user issues(2) Investigating issues using large amounts of text documents

(a) Identifying relevant factors (such as environment or cost), here referred to as “aspects”

(b) Retrieving grounds or counterexamples(3) Providing argument scripts in a persuasive manner

The key to the debating AI is the quality of grounds and counterexamples retrieved from large amounts

of text data. This section describes the evidence recognition technique used to retrieve this information (2-b).

In this step, the AI recognizes whether a retrieved sentence represents evidence for a predefined claim or not. The claim consists of an argument topic and an argument aspect. For example, that “electric cars” are good for “the environment” is a claim. Evidence comes in two types. Positive evidence is referred to as “grounds” and negative evidence as “counterexamples.” While grounds support the claim, counterexamples argue against it.

In trying to find counterexamples, it is important to accurately recognize whether a sentence is a ground or a counterexample. When composing arguments about electric cars, for example, a traditional text search would find a lot of positive evidence such as their producing zero emissions, which makes them good for the environment. On the other hand, a small amount of negative evidence, such as a manufacturing process imposing a large burden on the environment, is likely to be overlooked. However, these counterexamples are important because they can have a significant influence on decision-making. In order to find this important but infrequent evidence, it is necessary to recognize whether a sentence contains positive or negative evidence.

The recognition technique used for this purpose is as follows. In practical situations, many sentences have a more complicated syntax structure than, say, “electric cars are good for the environment.” Therefore, the authors constructed a technique to interpret sentences by folding partial linguistic structures into two hierarchical feature structures. The recognition steps are as follows.(1) Extract partial relations between contextual words as the first feature of the sentence(2) Combine them into relations of relations as the second feature of the sentence(3) Calculate a score for the sentence using machine learning

In the example shown in Fig. 3, step (1) extracts five partial linguistic structures, such as whether a subject of a causal verb refers to the topic or the aspect, or whether a specific word is present that reverses the meaning (positive or negative) of the text. In step (3), machine learning can achieve more accurate classification in a broader range of cases than manually created rules.

The debating AI achieved 77-point accuracy in evidence recognition. Future work will include

154 Debating Artifi cial Intelligence

- 76 -

improving the accuracy by using more complicated models of machine learning.

SOCIAL IMPLEMENTATION THROUGH OPEN COLLABORATIVE CREATION

The goal of the debating AI is to automatically investigate business opportunities and risks on a regular basis by integrating various information resources. To implement the vision, it is necessary to combine theories of corporate strategy, managers’ experiences, and a novel style of providing textual content. For example, it is possible that the form of news articles might change in the future. While news articles currently assume humans readers, in the future AIs may read news articles to support enterprise decision-making. Hitachi plans to start open collaborative creation in which various experts collaborate with one another.

CONCLUSIONS

This article has described an overview of the debating AI. Society has high expectations for AI technology. However, the actual level of the technology has yet to reach these expectations. To implement the vision, Hitachi will continue to improve the accuracy of the

arguments that the debating AI provides. Hitachi will also work to build open collaborative creation relationships.

ACKNOWLEDGMENTSThe authors would like to thank Prof. Kentaro Inui from Tohoku University for his valuable discussions.

REFERENCES(1) R. G. McGrath, “The End of Competitive Advantage: How

to Keep Your Strategy Moving as Fast as Your Business,” Harvard Business School Press (2013).

(2) M. Sato et al., “End-to-end Argument Generation System in Debating,”ACL-IJCNLP (2015).

(3) T. Yanase et al., “Learning Sentence Ordering for Opinion Generation of Debate,” 2nd Workshop on Argumentation Mining, (2015).

Counterexample or not

Machine learning

“Relation of relations” feature

Relation feature

no X X

No research showed that casino ban suppressed economy .

show X ban X X suppress suppress

Fig. 3—Evidence Recognition Technique.The technique extracts linguistic features from a given text. The technique recognizes whether a sentence is evidence or not by extracting information about causal effects, positive and negative factors, and sources, and then calculates a score using a machine learning technique.


- 77 -

Yoshiyuki Kobayashi, Ph.D.

Intelligent Information Research Department, Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi. Ltd. He is currently engaged in research into natural language processing. Dr. Kobayashi is a member of JSAI, the Information Processing Society of Japan (IPSJ), The Association for Natural Language Processing (Japan) (ANLP), and the Association for Computing Machinery (ACM).

Toshihiko Yanase, Ph.D.

Center for Exploratory Research, Research & Development Group, Hitachi. Ltd. He is currently engaged in research into debating artificial intelligence. Dr. Yanase is a member of JSAI, and the ACM.

Yoshiki Niwa, Ph.D.

Center for Exploratory Research, Research & Development Group, Hitachi. Ltd. He is currently engaged in research into natural language processing and text search. Dr. Niwa is a member of JSAI, IPSJ, and ANLP.

Kohsuke Yanai, Ph.D.

Center for Exploratory Research, Research & Development Group, Hitachi. Ltd. He is currently engaged in research of debating artificial intelligence. Dr. Yanai is a member of The Japanese Society for Artificial Intelligence (JSAI).

Misa Sato

Center for Exploratory Research, Research & Development Group, Hitachi. Ltd. She is currently engaged in research into debating artificial intelligence. Ms. Sato is a member of JSAI.

Toshinori Miyoshi, Ph.D.

Center for Exploratory Research, Research & Development Group, Hitachi. Ltd. He is currently engaged in research into natural language processing and brain science. Dr. Miyoshi is a member of JSAI, the Institute of Electronics, Information and Communication Engineers (IEICE), and the IEEE.

Hisashi Ikeda, Ph.D.

Center for Technology Innovation – Systems Engineering, Research & Development Group, Hitachi. Ltd. He is currently engaged in the management of systems research. Dr. Ikeda is a member of JSAI, the IPSJ, the IEICE, the IEEE, and the ACM.

ABOUT THE AUTHORS


- 78 -

Featured Articles

Advanced Research into AI

Ising Computer

Masanao Yamaoka, Ph.D.

Chihiro Yoshimura

Masato Hayashi

Takuya Okuyama

Hidetaka Aoki

Hiroyuki Mizuno, Ph.D.

OVERVIEW: A major challenge facing AI is the enormous computational load it imposes, of which combinational optimization makes up a large part. Hitachi has devised a computing technology based on a new paradigm that is capable of solving combinatorial optimization problems efficiently using an Ising model, and has built a prototype 20k-spin Ising computer chip using a 65-nm process. An Ising chip represents a combinatorial optimization problem by mapping it onto an Ising model based on the spin of magnetic materials, and solves the problem by taking advantage of the system’s natural tendency to converge. This convergence is implemented using a CMOS circuit. In addition to demonstrating its ability to solve combinatorial optimization problems and operate at 100 MHz, the prototype chip has been demonstrated to consume approximately 1,800 times less power to obtain the solution than would be required by a conventional computer with a von Neumann architecture.

INTRODUCTION

A major challenge facing artificial intelligence (AI) is the enormous computational load it imposes. This is because, in contrast to the conventional practice of mechanical execution of an algorithm defined by hand in the form of a program, AI learns automatically from data and uses this as the basis for realtime decision-making. Combinational optimization forms a large part of the heavy processing load associated with both the learning and decision-making steps. When an AI learns from data, for example, it needs to optimize the model parameters in order to minimize error. Similarly, when subsequently using the model for decision-making, the AI needs to optimize the decision parameters in order to maximize a performance function. In both cases, this combinational optimization requires finding the parameters that best satisfy the conditions out of a large number of possibilities, a problem that is difficult to solve efficiently using conventional computing practices.

Accordingly, Hitachi has developed a new concept in computing that can efficiently solve combinatorial optimization problems by using an Ising model, a statistical mechanics model that mimics the behavior of a magnetic material. Tests conducted on a prototype demonstrated that combinational optimization

problems could be solved with an efficiency three orders of magnitude or greater compared to conventional computing practices. This article describes this Ising computer.

COMBINATIONAL OPTIMIZATION PROBLEMS

A combinatorial optimization problem involves finding a solution that maximizes (or minimizes) a performance index under given conditions. A characteristic of combinatorial optimization problems is that the number of candidate solutions increases explosively the greater the number of parameters that define the problem. As the number of parameters in AI computation is increasing, the number of candidate solutions to combinational optimization problems is expected to increase dramatically in the future.

The solution of combinatorial optimization problems using existing computing techniques involves calculating the performance index for all parameter combinations and then selecting the combination that results in the minimum performance index [see Fig. 1 (a)]. Because the number of combinations for a problem with n parameters is 2n, a 1,000 parameter problem requires the performance indices to be calculated for 21000 parameter combinations (roughly


- 79 -

10300). Calculating performance indices for such a huge number of combinations is impossible in practice.

What is actually done in situations like this is that, rather than calculating the performance indices for all combinations, an approximation algorithm is used to obtain a roughly optimal combination of parameters. Unfortunately, as the number of parameters increases, finding even an approximate solution becomes difficult. Furthermore, semiconductor scaling has enabled the computational methods used in the past to deal with larger problems by improving the performance of the central processing units (CPUs) used for the calculations. However, progress on semiconductor scaling appears to have plateaued in recent years, and in practice there have been no further improvements in CPU clock speeds since the late 2000s. In other words, optimizing the larger and more complex systems of the future will require new computing techniques that do not rely on the practices of the past.

NEW COMPUTING CONCEPT

Conventional computers break problems down into a collection of programs (procedures) and solve the problems by executing these programs sequentially. As noted above, however, the difficulty with solving combinatorial optimization problems is the explosive growth in the number of procedures required for program execution. Accordingly, Hitachi has proposed adopting a different computing concept, namely natural computing.

Fig. 1 (b) shows the calculation procedure using natural computing. Natural computing works by using a natural phenomenon to model the problem to be solved (mapping) and taking advantage of the convergence inherent in this natural phenomenon to converge on the solution to the problem. The problem can then be solved by observing this converged result.

An Ising model, meanwhile, represents the behavior of magnetic spin in a magnetic material in terms of statistical mechanics and has been proposed as a suitable technique for solving combinatorial optimization problems. Fig. 2 shows an Ising model. The properties of a magnetic material are determined by magnetic spins, which can be oriented up or down. An Ising model is expressed in terms of the individual spin states (σi), the interaction coefficients (Jij) that represent the strength of the interactions between different pairs of spin states, and the external magnetic coefficients (hi) that represent the strength of the external magnetic field. The figure also includes the equation for the energy (H) of the Ising model. One property of an Ising model is that the spins shift to the states that minimize this energy, ultimately leaving the model in this minimum state. If a combinatorial optimization problem is mapped onto an Ising model in such a way that its performance index corresponds to the model’s energy, the Ising model is allowed to converge so that the spin states adopt the minimum-energy configuration. This is equivalent to obtaining the combination of parameters that minimizes the performance index of the original optimization problem.

(a) Conventional computing (b) Natural computing

Problem

Model problem using a natural phenomenon (mapping)

Convergence of natural phenomenon

Problem

SolutionSolution

2 timesn

Observe resultAssess indices

Vary parameters and recalculate performance indices

Fig. 1—Procedure for Solving Optimization Problems.Conventional practice has been to repeatedly calculate all performance indices and assess the values obtained. Natural computing, in contrast, reduces the number of calculation iterations by taking advantage of the tendency for a natural system to converge.

Σ Σ= − −σH Jij ii j i,

σhi iσj

J 78 J 89

J 58J 47

J 45 J 56

J 23J 12

J 25J 14

J 69

J 36

Fig. 2—Ising Model.An Ising model represents the properties of ferromagnetic materials in terms of statistical mechanics. It consists of a lattice of points (spins), each of which can occupy one of two orientation states, and reaches stability when the energy H is at a minimum, taking account of interactions between adjacent points in the lattice.

158 Ising Computer

- 80 -

CMOS ISING COMPUTING

While computing methods that use superconductors to

replicate an Ising model have been proposed in the past,

Hitachi has proposed using a complementary metal

oxide semiconductor (CMOS) circuit for this purpose.

The benefits of using a CMOS circuit are simpler

manufacturing, greater scalability, and ease of use.

The updating of actual spin values is performed in

accordance with the following rule:

New spin value = +1 (if a > b)

−1 (if a < b)

+/−1 (if a = b)

Here, a is the number of cases in which (adjacent

spin value, interaction coefficient) is (+1, +1) or

(−1, −1) and b is the number of cases in which it

is (+1, −1) or (−1, +1). These interactions cause

the energy of the Ising model to fall, following the

energy contours (landscape) like that shown in Fig. 3.

However, because the energy profile includes peaks

and valleys (as shown in the figure), this interaction

process operating on its own has the potential to leave

the model trapped in a local minimum in a region that

is not the overall minimum for the system.

To escape such local minima, the spin states

are randomly perturbed. This causes the system to

randomly switch to an unrelated state, as indicated

by the dotted line in Fig. 3. Collectively, these two

processes are called CMOS annealing. By using them,

it is possible to identify the state with the lowest energy

that can be found.

In practice, this use of random numbers means that

the solution obtained is not necessarily the optimal

one. However, when the computing technique is used

for parameter optimization, it is likely that it will not

matter if the results obtained are not always optimal.

In situations where this computing technique might

be deployed, it is possible to anticipate applications

where providing a theoretical guarantee that it will

produce solutions with 99% or better accuracy, 90%

or more of the time, for example, will mean that these

solutions can be relied on to not cause any problems

for the system.

PROTOTYPE COMPUTER

A prototype Ising chip was manufactured using a 65-nm

CMOS process to test the proposed Ising computing

technique. An Ising node was then built with this Ising

chip and its ability to solve optimization problems was

demonstrated. This section describes the prototype and

the results of its use to solve optimization problems.

CMOS Ising ChipThe prototype Ising chip was fabricated using a

65-nm semiconductor CMOS process. Fig. 4 shows

a photograph of the chip. The 3-mm × 4-mm chip

can hold 20,000 spin circuits, each occupying an area

of 11.27 μm × 23.94 μm ≈ 270 μm2. The interface

circuit used for reading and writing the spin states

and interaction coefficients operates at 100 MHz, as

does the interaction process for updating spin values.

Spin state (2n combinations)

Ener

gy H

(per

form

ance

index

)

Optimal solution

n: Number of spins

Fig. 3—Ising Model Energy Landscape and CMOS Annealing.In Ising computing, although the energy falls in accordance with the energy contours (landscape) due to the interactions between spins (solid arrows), there is a potential for it to get trapped at a local minimum. This can be prevented by inputting random numbers to deliberately invert spin values (dotted arrows). Called CMOS annealing, this operation obtains a solution with low energy.

1k-spin sub-array780×380 μm2

4 mm

3 m

m

1k-spin

SRAM I/F

Fig. 4—Ising Chip Photograph.The chip has 20,000 spin circuits in a 3-mm × 4-mm = 12-mm2 area.

SRAM: static random access memory I/F: interface


- 81 -

The Ising chip implements a three-dimensional

Ising model on a two-dimensional memory lattice.

Semiconductor chips achieve a high level of integration

by using a two-dimensional layout, and the prototype

Ising chip also takes advantage of this to achieve a high

level of integration, meaning that it can implement a

large number of spin circuits.

Ising ComputerFig. 5 shows a prototype Ising node fitted with two

Ising chips.

The Ising node can be accessed from a personal

computer (PC) or server via a local area network

(LAN) to input optimization problems and obtain the

solutions.

Fig. 6 shows a comparison of the energies required

to solve a randomly generated maximum cut problem

using the Ising node and using conventional computing.

The horizontal axis represents the number of spins

in the Ising model. The conventional computing

technique used for comparison consisted of executing

the SG3 approximation algorithm (which has been

optimized for solving maximum cut problems) on a

general-purpose CPU. The same problems were solved

using both techniques and a comparison was made of

the energies consumed in obtaining a solution to an

equivalent level of accuracy in each case. Because the

SG3 approximation algorithm used for the comparison

had been optimized for maximum cut problems that

use Ising models, there was no significant difference

between the times taken by the two techniques for a

problem with 20,000 spins. The amount of energy

consumed in solving a 20,000-spin problem, however,

was approximately 1,800 times less using the new

technique.

CONCLUSIONS

Table 1 shows a comparison with previous Ising

computers. Use of a CMOS semiconductor circuit

means the computer can operate at room temperature.

This means low power consumption for cooling

is achieved. While the prototype computer has

approximately 20,000 spin circuits, it will be possible

to replicate even larger Ising models by using higher

levels of semiconductor process scaling.

Furthermore, because the current system uses

digital values to calculate spin interactions, it is easy

to link a number of chips together and expand the size

by using multiple chips.

Although it is anticipated that this use of digital

circuits will result in lower solution accuracy

Ising chips

Fig. 5—Ising Node.The photograph shows an Ising node with two Ising chips. The Ising node is connected to a server or PC via a LAN cable and can be used to solve combinatorial optimization problems.

PC: personal computer LAN: local area network

81

10

100

1,000

10,000

64 512

Number of spins (problem size)

Ener

gy e

ffic

iency

(rel

ativ

e to

conven

tional

appro

xim

atio

n a

lgori

thm

)

4,096

×1,800

32,768

Fig. 6—Energy Efficiency of Solving Randomly Generated Maximum Cut Problem.The graph shows the relative energy efficiency of the calculation compared to an approximation algorithm executing on a general-purpose CPU. The energy efficiency improves as the size (number of spins) of the problem increases, with the new technique being approximately 1,800 times more efficient for a 20,000-spin problem.

New technique Existing technique

ApproachIsing computing

Semiconductor (CMOS) Superconductor

Operating temperature

Room temperature 20 mK

Power consumption

0.05 W15,000 W (including

cooling)

Scalability (number of spins)

20,000 (65 nm)Can be scaled up by using higher level of

scaling

512

Computation time

MillisecondsMilliseconds (fast in

principle)

TABLE 1. Comparison with Existing Ising Computers

The new technique is significant in engineering terms because of its suitability for real-world applications, being superior to an existing Ising computer that uses superconductors in terms of things like ease-of-use and scalability.

160 Ising Computer

- 82 -

optimization problems in AI applications, which are

expected to impose increasing processing loads in

the future.

REFERENCES(1) M. W. Johnson et al., “Quantum Annealing with Manufactured

Spins,” Nature 473, pp. 194–198 (May 2011).

(2) R. F. Service, “The Brain Chip,” Science 345, Issue 6197

(Aug. 2014).

(3) C. Yoshimura et al., “Spatial Computing Architecture

Using Randomness of Memory Cell Stability under Voltage

Control,” 2013 European Conference on Circuit Theory and

Design (Sep. 2013).

(4) M. Yamaoka et al., “20k-spin Ising Chip for Combinational

Optimization Problem with CMOS Annealing,” ISSCC 2015

digest of technical papers, pp. 432–433 (Feb. 2015).

(5) S. Kahruman et al., “On Greedy Construction Heuristics

for the MAX-CUT Problem,” International Journal of

Computational Science and Engineering 3, No. 3, pp. 211–

218 (2007).

than can be achieved by previous systems based

on superconductors, it is adequate for use in the

optimization of actual social systems because it is able

to solve problems in practice. Moreover, the approach

described here of using a semiconductor is significant

in engineering terms for reasons that include ease-of-

use and scalability.

This article has described how the prototype Ising

computer successfully solved a maximum cut problem,

which is a form of combinatorial optimization

problem. As it is known that this problem can be

translated mathematically into other combinatorial

optimization problems, this indicates that the technique

has potential for use in actual system optimization.

Furthermore, energy measurements demonstrated that

the technique can reduce consumption by three or

more orders of magnitude compared to conventional

computing techniques.

In the future, Hitachi sees the Ising computer as a

highly efficient technology for solving combinational

Chihiro Yoshimura

Center for Exploratory Research, Research & Development Group, Hitachi, Ltd. He is currently engaged in research into computers based on new concepts.

Takuya Okuyama


Hiroyuki Mizuno, Ph.D.

Center for Technology Innovation–Information and Telecommunications, Research & Development Group, Hitachi, Ltd. He is currently engaged in research into information and telecommunications technology. Dr. Mizuno is a member of the IEEE.

Masanao Yamaoka, Ph.D.

Center for Exploratory Research, Research & Development Group, Hitachi, Ltd. He is currently engaged in research into computers based on new concepts. Dr. Yamaoka is a member of the IEEE.

Masato Hayashi


Hidetaka Aoki

Hitachi Asia (Malaysia) Sdn. Bhd. He is currently engaged in the research and development of green computing. Mr. Aoki is a member of the Information Processing Society of Japan.

ABOUT THE AUTHORS

Planning Committee

Norihiro Suzuki (chairman)Nobuya AbematsuKunio UchiyamaKenji KatouKeiichiro NakanishiTakashi HottaKazuo MinamiKazuaki OtomoMasayuki ShimonoTakeshi YoshikawaMasahiro MimuraYasushi YokosukaKouji NomuraTakahiro TachiYoshiki KakumotoShuuichi KannoTakayuki SuzukiTakeshi InoueAkira Banno

Hitachi Review Volume 65 Number 6 July 2016

ISSN 0018-277XHitachi Review is published by Hitachi, Ltd.Visit our site at www.hitachi.com/revAddress correspondence to: The Editor, Hitachi Review, Advertising Dept., Corporate Brand & Communications Div., Hitachi, Ltd.Shin-Otemachi Building, 2-1, Otemachi 2-chome, Chiyoda-ku, Tokyo, 100-0004 JapanEditor-in-Chief: Akira Banno©2016 Hitachi, Ltd.Date of Issue: July, 2016Printed in Japan by Hitachi Document Solutions Co., Ltd.

XR-E064

This Issue’s Editorial Coordinators

Norihiro SuzukiKazuo Yano

Artiﬁcial Intelligence as a Hope

Documents

Transcript of Artiﬁcial Intelligence as a Hope